{"id":90138,"date":"2026-08-04T19:09:54","date_gmt":"2026-08-04T11:09:54","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/90138.html"},"modified":"2026-08-04T19:09:54","modified_gmt":"2026-08-04T11:09:54","slug":"ai4eda-ai4design-%e5%ae%8c%e6%95%b4%e7%a0%94%e7%a9%b6%e6%8a%a5%e5%91%8a%ef%bc%9a%e5%85%a8%e7%90%83%e8%ae%ba%e6%96%87%e3%80%81%e9%a1%b9%e7%9b%ae%e3%80%81%e6%9c%ba%e6%9e%84%e4%b8%8e%e4%ba%a7%e4%b8%9a","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/90138.html","title":{"rendered":"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf"},"content":{"rendered":"<h2>AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a&#xff1a;\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf<\/h2>\n<p>\u539f\u6587\u5305\u542b\u5927\u91cf\u8bba\u6587\u622a\u56fe\u548c\u6765\u6e90\u94fe\u63a5\u3002\u5b8c\u6574\u7248\u89c1&#xff1a;https:\/\/ai4eda-ai4design-report-20260624-qcy.netlify.app<\/p>\n<p>\u516c\u4f17\u53f7 \u201c\u82af\u7247\u4eba-\u6652 AI \u7b14\u8bb0\u201d<\/p>\n<p>\u672c\u6587\u91c7\u7528\u201c\u4e00\u6761\u7eb5\u8f74\u3001\u4e94\u6761\u6a2a\u8f74\u201d\u7684\u7ed3\u6784&#xff1a;NVIDIA 2023-2026 \u5e74 19 \u7bc7\u8bba\u6587\u4f5c\u4e3a\u8fde\u7eed\u7eb5\u8f74&#xff0c;Google\/DeepMind\u3001DREAMPlace\u3001OpenROAD\u3001EDA Corpus\u3001CircuitNet\u3001ForgeEDA\u3001Chip-Chat\u3001ChipGPT\u3001RTLLM\u3001NSF\u3001Synopsys\u3001Cadence\u3001Siemens\u3001NVIDIA cuLitho \u4e0e TSMC fab AI \u7b49\u4f5c\u4e3a\u5168\u7403\u6a2a\u8f74\u3002\u76ee\u6807\u4e0d\u662f\u53ea\u4ecb\u7ecd\u67d0\u5bb6\u516c\u53f8&#xff0c;\u800c\u662f\u7cfb\u7edf\u68b3\u7406 AI \u6b63\u5728\u5982\u4f55\u91cd\u5851\u82af\u7247\u8bbe\u8ba1\u3001\u9a8c\u8bc1\u3001\u7269\u7406\u5b9e\u73b0\u3001\u5f00\u653e\u6570\u636e\u3001\u5546\u4e1a EDA \u5e73\u53f0\u3001\u5236\u9020\u8ba1\u7b97\u548c fab operations\u3002<\/p>\n<p>\u672c\u6587\u5d4c\u5165\u7684\u8bba\u6587\u56fe\u7247\u5747\u6765\u81ea\u672c\u5730 PDF \u7684 Figure\/Table \u5c40\u90e8\u88c1\u56fe&#xff0c;\u4e0d\u4f7f\u7528\u6574\u9875\u622a\u56fe\u3002\u5b83\u4eec\u4f5c\u4e3a\u201c\u539f\u6587\u8bc1\u636e\u5207\u7247\u201d&#xff0c;\u7528\u4e8e\u8f85\u52a9\u89e3\u91ca\u6bcf\u7bc7\u8bba\u6587\u7684\u5173\u952e\u65b9\u6cd5\u548c\u7ed3\u679c\u3002<\/p>\n<h3>\u4ece NVIDIA \u7eb5\u8f74\u5230\u5168\u7403\u6a2a\u8f74<\/h3>\n<p>\u672c\u7248\u628a NVIDIA \u9010\u7bc7\u6df1\u8bfb\u5347\u7ea7\u4e3a\u5b8c\u6574\u6a2a\u7eb5\u7814\u7a76&#xff1a;\u8bba\u6587\u5185 Figure\/Table \u5c40\u90e8\u622a\u56fe\u4fdd\u7559\u4e3a\u8bc1\u636e\u5207\u7247&#xff0c;\u975e NVIDIA \u8bba\u6587\u3001\u5f00\u6e90\u9879\u76ee\u3001\u673a\u6784\u62a5\u544a\u548c\u4ea7\u4e1a\u5e73\u53f0\u88ab\u63d0\u5347\u4e3a\u540c\u7b49\u91cd\u91cf\u7684\u4e3b\u7ae0\u8282\u3002<\/p>\n<h3>\u8bfb\u6cd5&#xff1a;\u4e00\u6761\u7eb5\u8f74&#xff0c;\u4e94\u6761\u6a2a\u8f74<\/h3>\n<p>\u672c\u62a5\u544a\u6309\u7528\u6237\u786e\u8ba4\u7684 A \u7ed3\u6784\u91cd\u7ec4&#xff1a;NVIDIA \u662f\u4e00\u6761\u8fde\u7eed\u8bba\u6587\u7eb5\u8f74&#xff0c;\u5168\u7403\u8bba\u6587\u3001\u5f00\u6e90\u9879\u76ee\u3001\u673a\u6784\u8def\u7ebf\u56fe\u548c\u4ea7\u4e1a\u5e73\u53f0\u6784\u6210\u6a2a\u5411\u5bf9\u7167\u3002<\/p>\n<h4>NVIDIA \u7eb5\u8f74<\/h4>\n<p>19 \u7bc7\u8bba\u6587\u4ece VerilogEval\u3001ChipNeMo\u3001RTLFixer \u8d70\u5230 ACE-RTL\u3001Trace2Skill&#xff0c;\u5c55\u793a LLM\/agent \u5982\u4f55\u4e00\u6b65\u6b65\u63a5\u5165\u7f16\u8bd1\u5668\u3001\u4eff\u771f\u5668\u3001\u5f62\u5f0f\u9a8c\u8bc1\u3001STA\u3001trace \u548c skill evolution\u3002<\/p>\n<h4>\u5b66\u672f\u6a2a\u8f74<\/h4>\n<p>AlphaChip\u3001DREAMPlace\u3001Chip-Chat\u3001ChipGPT\u3001RTLLM\u3001OpenLLM-RTL\u3001RTL-BenchLS \u7b49\u8bba\u6587\u5c55\u793a\u4e86 RL\/GNN\u3001GPU \u4f18\u5316\u3001\u81ea\u7136\u8bed\u8a00\u5230 RTL\u3001\u53ef\u9a8c\u8bc1 benchmark \u7684\u4e0d\u540c\u8def\u5f84\u3002<\/p>\n<h4>\u5f00\u653e\u751f\u6001\u6a2a\u8f74<\/h4>\n<p>OpenROAD\u3001EDA Corpus\u3001CircuitNet\u3001ForgeEDA \u628a AI4EDA \u4ece\u79c1\u6709\u5b9e\u9a8c\u63a8\u5411\u53ef\u590d\u73b0\u57fa\u7840\u8bbe\u65bd&#xff1a;\u5f00\u653e flow\u3001\u5f00\u653e\u811a\u672c\u6570\u636e\u3001\u5f00\u653e layout\/\u7f51\u8868\/\u591a\u6a21\u6001\u6570\u636e\u3002<\/p>\n<h4>\u4ea7\u4e1a\u6a2a\u8f74<\/h4>\n<p>Synopsys.ai\u3001Cadence Cerebrus\/ChipStack\/AgentStack\u3001Siemens Solido\/Aprisa \u548c NVIDIA cuLitho\/TSMC fab AI&#xff0c;\u5c55\u793a\u5546\u4e1a EDA \u5982\u4f55\u628a AI \u653e\u8fdb PPA\u3001\u9a8c\u8bc1\u3001\u6a21\u62df\u3001\u5236\u9020\u548c fab operations\u3002<\/p>\n<h4>\u8def\u7ebf\u56fe\u6a2a\u8f74<\/h4>\n<p>ML for EDA Survey\u3001LLM for EDA Survey\u3001NSF AI4EDA Workshop Report \u7ed9\u51fa\u5b66\u79d1\u5c42\u7ea7\u7684\u8fb9\u754c&#xff1a;\u6570\u636e\u3001\u7b97\u529b\u3001benchmark\u3001\u7b7e\u6838\u3001\u4eba\u624d\u548c\u8de8\u5b66\u79d1\u534f\u4f5c\u3002<\/p>\n<h4>\u5224\u65ad\u6a2a\u8f74<\/h4>\n<p>\u771f\u6b63\u53ef\u843d\u5730\u7684 AI4EDA&#xff0c;\u4e0d\u662f\u5355\u4e00\u6a21\u578b\u6216\u5355\u4e00\u8bba\u6587&#xff0c;\u800c\u662f\u201c\u7ed3\u6784\u5316\u8bbe\u8ba1\u5bf9\u8c61 &#043; \u5de5\u5177\u53cd\u9988 &#043; \u53ef\u9a8c\u8bc1\u6570\u636e &#043; agent \u8f68\u8ff9 &#043; \u4ea7\u4e1a flow\u201d\u7684\u7ec4\u5408\u3002<\/p>\n<h4>\u5168\u7403\u6a2a\u8f74\u5bf9\u8c61\u7d22\u5f15<\/h4>\n<p>Google DeepMind \u5f3a\u5316\u5b66\u4e60 &#043; GNN \u5b8f\u5355\u5143\u5e03\u5c40<\/p>\n<p>UT Austin \/ \u5f00\u6e90\u793e\u533a GPU \u52a0\u901f\u5206\u6790\u5f0f placement<\/p>\n<p>UCSD \/ DARPA \/ \u5f00\u6e90\u793e\u533a \u5f00\u653e RTL-to-GDS \u4e0e no-human-in-loop<\/p>\n<p>OpenROAD Assistant \u9762\u5411 OpenROAD \u7684 LLM QA\/\u811a\u672c\u6570\u636e<\/p>\n<p>\u5b66\u672f\u793e\u533a \u7269\u7406\u8bbe\u8ba1\u9884\u6d4b\u4efb\u52a1\u5f00\u653e\u6570\u636e\u96c6<\/p>\n<p>\u5b66\u672f\u793e\u533a \u591a\u6a21\u6001\u7535\u8def\u8868\u793a\u4e0e PPA \u4efb\u52a1\u6570\u636e\u96c6<\/p>\n<p>NYU \/ \u76f8\u5173\u5b66\u672f\u56e2\u961f \u5bf9\u8bdd\u5f0f\u786c\u4ef6\u8bbe\u8ba1\u4e0e AI-written HDL tapeout<\/p>\n<p>\u5b66\u672f\u56e2\u961f \u81ea\u7136\u8bed\u8a00\u5230\u786c\u4ef6\u903b\u8f91\u7684\u56db\u9636\u6bb5 zero-code flow<\/p>\n<p>\u5b66\u672f\u56e2\u961f \u81ea\u7136\u8bed\u8a00\u5230 RTL \u5f00\u653e benchmark<\/p>\n<p>\u5b66\u672f\u7efc\u8ff0 LLM \u524d\u7684 ML-for-EDA \u65b9\u6cd5\u7248\u56fe<\/p>\n<p>\u5b66\u672f\u7efc\u8ff0 LLM \u8fdb\u5165 EDA \u7684\u65b9\u6cd5\u5206\u7c7b<\/p>\n<p>NSF \/ NeurIPS Workshop \u6570\u636e\u3001\u7b97\u529b\u3001\u4eba\u624d\u4e0e\u534f\u4f5c\u8def\u7ebf\u56fe<\/p>\n<p>Synopsys AI-driven EDA \/ GenAI \/ Agentic AI \u5168\u6808\u5e73\u53f0<\/p>\n<p>Cadence PPA \u4f18\u5316\u3001\u9a8c\u8bc1 agent\u3001\u5de5\u7a0b super agent<\/p>\n<h3>\u6458\u8981\u4e0e\u603b\u8bba&#xff1a;NVIDIA \u4e0d\u662f\u5168\u90e8&#xff0c;\u5b83\u662f\u4e00\u6761\u6700\u6e05\u695a\u7684\u7eb5\u8f74<\/h3>\n<p>AI4EDA \u4e0e AI4Design \u7684\u6838\u5fc3\u53d8\u5316&#xff0c;\u4e0d\u662f\u201c\u8ba9\u5927\u6a21\u578b\u5199 Verilog\u201d\u8fd9\u4e48\u7b80\u5355&#xff0c;\u800c\u662f\u628a\u82af\u7247\u5de5\u7a0b\u4e2d\u7684\u89c4\u683c\u3001\u4ee3\u7801\u3001\u811a\u672c\u3001\u65e5\u5fd7\u3001\u6ce2\u5f62\u3001STA report\u3001formal collateral\u3001layout image\u3001PPA \u6307\u6807\u3001\u5236\u9020\u8ba1\u7b97\u7b49\u5bf9\u8c61\u7edf\u4e00\u7eb3\u5165\u53ef\u5b66\u4e60\u3001\u53ef\u68c0\u7d22\u3001\u53ef\u9a8c\u8bc1\u3001\u53ef\u8fed\u4ee3\u7684\u95ed\u73af\u3002NVIDIA \u7684\u8bba\u6587\u8c31\u7cfb\u6e05\u6670\u5448\u73b0\u4e86\u8fd9\u6761\u8def\u5f84&#xff1a;\u5148\u5efa\u7acb\u53ef\u6267\u884c\u8bc4\u6d4b&#xff0c;\u518d\u505a\u9886\u57df\u6a21\u578b\u548c\u6570\u636e\u5de5\u7a0b&#xff0c;\u7136\u540e\u8fdb\u5165\u5de5\u5177\u53cd\u9988\u3001\u591a\u667a\u80fd\u4f53\u3001\u957f\u4e0a\u4e0b\u6587\u3001verifier-guided skill evolution\u3002\u4e0e\u6b64\u540c\u65f6&#xff0c;\u5168\u7403\u4ea7\u4e1a\u754c\u628a\u540c\u4e00\u903b\u8f91\u843d\u5230 PPA\u3001TAT\u3001mask throughput \u548c\u5148\u8fdb\u5de5\u827a scaling \u4e0a\u3002<\/p>\n<p>VerilogEval\u3001FVEval\u3001CVDP \u628a\u8bc4\u6d4b\u4ece\u6587\u672c\u76f8\u4f3c\u8f6c\u5411\u7f16\u8bd1\u3001\u4eff\u771f\u3001formal\u3001hidden tests \u4e0e repository-level workflow\u3002\u8fd9\u662f AI4EDA \u6210\u719f\u7684\u7b2c\u4e00\u9053\u95e8\u69db\u3002<\/p>\n<p>ChipNeMo\u3001ChipAlign\u3001JARVIS \u8868\u660e\u82af\u7247\u77e5\u8bc6\u6765\u81ea\u79c1\u6709\u6587\u6863\u3001\u5de5\u5177 API\u3001RTL\u3001bug\u3001\u811a\u672c\u548c\u65e5\u5fd7&#xff1b;RAG\u3001tokenizer\u3001\u9886\u57df\u7ee7\u7eed\u9884\u8bad\u7ec3\u548c\u6307\u4ee4\u5bf9\u9f50\u5fc5\u987b\u7ec4\u5408\u4f7f\u7528\u3002<\/p>\n<p>RTLFixer\u3001VerilogCoder\u3001Timing Agent\u3001PRO-V-R1\u3001ACE-RTL\u3001Trace2Skill \u90fd\u628a\u7f16\u8bd1\u5668\u3001\u4eff\u771f\u5668\u3001STA\u3001formal\u3001verifier feedback \u653e\u8fdb agent \u5faa\u73af\u3002<\/p>\n<p>CVDP\u3001ACE-RTL\u3001Trace2Skill \u4ee3\u8868\u65b0\u9636\u6bb5&#xff1a;\u771f\u6b63\u96be\u70b9\u662f\u591a\u6587\u4ef6\u4f9d\u8d56\u3001reuse\u3001\u9a8c\u8bc1\u3001debug\u3001\u73af\u5883\u914d\u7f6e\u548c\u5931\u8d25\u8f68\u8ff9\u6c89\u6dc0\u3002<\/p>\n<p>Multimodal PD Assistant\u3001CircuitNet\u3001ForgeEDA \u8bf4\u660e\u540e\u7aef\u8bbe\u8ba1\u7684\u5173\u952e\u8bc1\u636e\u662f\u56fe\u50cf\u3001\u70ed\u56fe\u3001\u8868\u683c\u3001\u7f51\u8868\u3001\u56fe\u7ed3\u6784\u548c\u62a5\u544a&#xff0c;\u800c\u4e0d\u662f\u7eaf\u81ea\u7136\u8bed\u8a00\u3002<\/p>\n<p>AlphaChip\u3001DREAMPlace\u3001Synopsys.ai\u3001Cadence Cerebrus\u3001cuLitho \u5171\u540c\u8868\u660e\u4ea7\u4e1a\u76ee\u6807\u4e0d\u662f pass&#064;k&#xff0c;\u800c\u662f PPA\u3001\u5468\u8f6c\u5468\u671f\u3001\u7b97\u529b\u6210\u672c\u3001mask throughput \u548c silicon scaling\u3002<\/p>\n<h3>\u53d1\u5c55\u8109\u7edc\u4e0e\u4e3b\u8981\u8fdb\u6b65\u70b9<\/h3>\n<h4>2019-2022&#xff1a;AI \u4f18\u5316\u5668\u4e0e GPU \u52a0\u901f EDA \u5148\u884c<\/h4>\n<p>DREAMPlace\u3001AlphaChip\u3001OpenROAD \u7b49\u5de5\u4f5c\u628a AI\/ML\/GPU \u7528\u4e8e placement\u3001autonomous flow \u548c\u5f00\u6e90\u8bbe\u8ba1\u81ea\u52a8\u5316&#xff0c;\u8bc1\u660e EDA \u53ef\u4ee5\u4ece\u4e13\u5bb6\u8c03\u53c2\u8f6c\u5411\u641c\u7d22\u3001\u5b66\u4e60\u548c\u52a0\u901f\u8ba1\u7b97\u3002<\/p>\n<h4>2023&#xff1a;LLM \u8fdb\u5165 RTL \u4e0e\u82af\u7247\u5de5\u7a0b\u8bed\u5883<\/h4>\n<p>VerilogEval\u3001ChipNeMo\u3001RTLFixer \u5efa\u7acb\u4e86 NVIDIA \u8def\u7ebf\u7684\u4e09\u5757\u57fa\u77f3&#xff1a;\u53ef\u6267\u884c\u8bc4\u6d4b\u3001\u9886\u57df\u6a21\u578b\u9002\u914d\u3001\u7f16\u8bd1\u53cd\u9988\u4fee\u590d\u3002<\/p>\n<h4>2024&#xff1a;\u4ece benchmark \u6269\u5c55\u5230 DSL\u3001formal\u3001agent \u4e0e\u6570\u636e\u5de5\u7a0b<\/h4>\n<p>PyHDL-Eval\u3001CraftRTL\u3001VerilogCoder\u3001FVEval\u3001ChipAlign \u8bf4\u660e\u95ee\u9898\u4e0d\u518d\u53ea\u662f\u751f\u6210 RTL&#xff0c;\u800c\u662f\u8986\u76d6\u591a\u8bed\u8a00\u786c\u4ef6\u8868\u8fbe\u3001\u5f62\u5f0f\u9a8c\u8bc1\u3001\u529f\u80fd\u8c03\u8bd5\u548c instruction alignment\u3002<\/p>\n<h4>2025&#xff1a;\u591a\u667a\u80fd\u4f53\u548c\u5de5\u7a0b\u4efb\u52a1\u6210\u4e3a\u4e3b\u7ebf<\/h4>\n<p>AssertionForge\u3001Marco\u3001Timing Agent\u3001JARVIS\u3001ScaleRTL\u3001PRO-V-R1\u3001CVDP\u3001Multimodal PD Assistant \u628a\u4efb\u52a1\u63a8\u5411 verification\u3001STA\u3001script\u3001long-context benchmark \u548c\u591a\u6a21\u6001\u7269\u7406\u8bbe\u8ba1\u3002<\/p>\n<h4>2026&#xff1a;\u4e0a\u4e0b\u6587\u548c\u6280\u80fd\u81ea\u8fdb\u5316<\/h4>\n<p>ACE-RTL \u4e0e Trace2Skill \u8868\u660e agent \u7684\u80fd\u529b\u4e0d\u53ea\u6765\u81ea\u6a21\u578b\u6743\u91cd&#xff0c;\u800c\u6765\u81ea\u4e0a\u4e0b\u6587\u6f14\u5316\u3001\u5931\u8d25\u8f68\u8ff9\u3001dense verifier feedback \u548c\u53ef\u5ba1\u8ba1 skill \u5e93\u3002<\/p>\n<h3>NVIDIA \u7eb5\u8f74&#xff1a;19 \u7bc7\u8bba\u6587\u9010\u7bc7\u6df1\u8bfb&#xff0c;\u4ece\u8bc4\u6d4b\u57fa\u5ea7\u5230\u81ea\u8fdb\u5316\u4ee3\u7406<\/h3>\n<h4>VerilogEval: Evaluating Large Language Models for Verilog Code Generation<\/h4>\n<p>2023 \u00b7 ICCAD 2023 \/ arXiv \u00b7 AI4Design \u00b7 Benchmark \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110947-6a71c87b39e25.png\" alt=\"Paper 01 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.6&#xff5c;Fig. 8 SFT training epochs and pass rate on VerilogEval. Dashed lines are gpt-3.5 results.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110947-6a71c87b95444.png\" alt=\"Paper 01 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.6&#xff5c;Detected table region by pdfplumber fallback, page 6, table 1, rows 5.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u628a Verilog \u751f\u6210\u4ece\u6587\u672c\u76f8\u4f3c\u5ea6\u95ee\u9898\u6539\u9020\u6210\u53ef\u7f16\u8bd1\u3001\u53ef\u4eff\u771f\u3001\u53ef\u590d\u73b0\u7684\u529f\u80fd\u6b63\u786e\u6027\u8bc4\u6d4b\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u5728\u8fd9\u7bc7\u8bba\u6587\u4e4b\u524d&#xff0c;\u5f88\u591a RTL \u751f\u6210\u5de5\u4f5c\u5bb9\u6613\u505c\u7559\u5728\u201c\u4ee3\u7801\u770b\u8d77\u6765\u50cf Verilog\u201d\u6216\u5c11\u91cf\u6837\u4f8b\u6f14\u793a\u4e0a\u3002\u786c\u4ef6\u8bbe\u8ba1\u7684\u6838\u5fc3\u4e0d\u662f\u8bed\u6cd5\u76f8\u4f3c&#xff0c;\u800c\u662f\u7efc\u5408\/\u4eff\u771f\u540e\u7684\u65f6\u5e8f\u884c\u4e3a\u662f\u5426\u5339\u914d\u89c4\u683c&#xff1b;\u5982\u679c\u6ca1\u6709\u81ea\u52a8\u5316 testbench \u548c golden transient output&#xff0c;LLM \u7684\u8fdb\u6b65\u65e0\u6cd5\u88ab\u7a33\u5b9a\u6bd4\u8f83\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u4ece HDLBits \u4e2d\u6574\u7406 156 \u4e2a\u95ee\u9898&#xff0c;\u8986\u76d6\u7ec4\u5408\u903b\u8f91\u3001\u65f6\u5e8f\u903b\u8f91\u548c FSM \u7b49\u4e0d\u540c\u96be\u5ea6\u3002<\/li>\n<li>\u7528 Icarus Verilog \u7b49\u5de5\u5177\u8fd0\u884c\u751f\u6210\u4ee3\u7801&#xff0c;\u5e76\u5c06\u77ac\u6001\u4eff\u771f\u8f93\u51fa\u4e0e golden solution \u6bd4\u8f83\u3002<\/li>\n<li>\u5b9a\u4e49 VerilogEval-Human \u548c VerilogEval-Machine \u4e24\u7c7b\u63d0\u793a\u8bbe\u7f6e&#xff0c;\u5e76\u7528 pass&#064;k \u8861\u91cf functional correctness\u3002<\/li>\n<li>\u5c1d\u8bd5\u7528 LLM \u751f\u6210\u7684 synthetic problem-code pairs \u505a\u76d1\u7763\u5fae\u8c03&#xff0c;\u9a8c\u8bc1\u9886\u57df\u6570\u636e\u5bf9 Verilog \u80fd\u529b\u7684\u63d0\u5347\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u786c\u4ef6\u4ee3\u7801\u8bc4\u6d4b\u7684\u57fa\u672c\u5355\u4f4d\u5e94\u662f\u884c\u4e3a\u7b49\u4ef7&#xff0c;\u800c\u4e0d\u662f\u6587\u672c\u76f8\u4f3c\u3002<\/li>\n<li>pass&#064;k \u53ea\u6709\u5728 testbench \u548c golden behavior \u8db3\u591f\u53ef\u9760\u65f6\u624d\u6709\u610f\u4e49\u3002<\/li>\n<li>Verilog \u751f\u6210\u7684\u7814\u7a76\u8d77\u70b9\u4e0d\u662f\u6a21\u578b&#xff0c;\u800c\u662f\u53ef\u590d\u73b0\u6c99\u7bb1\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u7528 HDLBits \u5f62\u6210 156 \u9898\u53ef\u6267\u884c benchmark\u3002<\/li>\n<li>\u628a transient simulation output \u4f5c\u4e3a\u529f\u80fd\u6b63\u786e\u6027\u5224\u65ad\u4f9d\u636e\u3002<\/li>\n<li>\u7528 LLM synthetic problem-code pairs \u63a2\u7d22\u4f4e\u6210\u672c SFT\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u660e\u786e\u6307\u51fa BLEU \u7c7b\u6587\u672c\u6307\u6807\u4e0d\u80fd\u533a\u5206\u786c\u4ef6\u884c\u4e3a\u662f\u5426\u6b63\u786e\u3002<\/li>\n<li>\u5b9e\u9a8c\u6bd4\u8f83\u4e86 human prompt \u4e0e machine prompt&#xff0c;\u5e76\u62a5\u544a SFT \u80fd\u63d0\u5347\u90e8\u5206\u6a21\u578b\u8868\u73b0\u3002<\/li>\n<li>\u540e\u7eed RTLFixer\u3001CraftRTL\u3001ScaleRTL\u3001CVDP \u5747\u6cbf\u7528\u5176\u53ef\u6267\u884c\u8bc4\u6d4b\u601d\u60f3\u3002<\/li>\n<li>\u9898\u76ee\u591a\u4e3a\u77ed\u6a21\u5757&#xff0c;\u4e0d\u80fd\u4ee3\u8868\u771f\u5b9e SoC repository \u590d\u6742\u5ea6\u3002<\/li>\n<li>\u6b63\u786e\u6027\u53d7 testbench \u8986\u76d6\u5f71\u54cd&#xff0c;\u672a\u8986\u76d6\u884c\u4e3a\u4ecd\u53ef\u80fd\u9519\u8bef\u3002<\/li>\n<li>\u4e0d\u76f4\u63a5\u5904\u7406\u65f6\u5e8f\u7ea6\u675f\u3001CDC\/RDC\u3001\u4f4e\u529f\u8017\u548c\u591a\u6587\u4ef6\u4f9d\u8d56\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; \u8fd9\u7bc7\u8bba\u6587\u7684\u4ef7\u503c\u4e0d\u5728\u4e8e\u63d0\u51fa\u66f4\u5f3a\u7684\u6a21\u578b&#xff0c;\u800c\u5728\u4e8e\u7ed9 NVIDIA \u540e\u7eed\u6240\u6709 RTL\/EDA LLM \u5de5\u4f5c\u63d0\u4f9b\u4e86\u5171\u540c\u8ba1\u91cf\u5355\u4f4d\u3002\u5b83\u660e\u786e\u5426\u5b9a\u4e86 BLEU\u3001\u7f16\u8f91\u8ddd\u79bb\u8fd9\u7c7b\u6587\u672c\u6307\u6807\u5728\u786c\u4ef6\u4ee3\u7801\u4e0a\u7684\u5145\u5206\u6027&#xff1a;\u4e24\u4e2a Verilog \u5b9e\u73b0\u53ef\u4ee5\u6587\u672c\u5dee\u5f02\u5f88\u5927\u4f46\u884c\u4e3a\u4e00\u81f4&#xff0c;\u4e5f\u53ef\u4ee5\u6587\u672c\u5f88\u50cf\u4f46\u529f\u80fd\u9519\u8bef\u3002VerilogEval \u628a\u95ee\u9898\u951a\u5b9a\u5230\u7f16\u8bd1\u5668\u548c\u4eff\u771f\u5668&#xff0c;\u8fd9\u5c31\u662f\u540e\u7eed RTLFixer\u3001CraftRTL\u3001ScaleRTL\u3001CVDP \u80fd\u6301\u7eed\u8fed\u4ee3\u7684\u539f\u56e0\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u5185\u90e8\u5efa\u7acb RTL assistant \u7684\u7b2c\u4e00\u5c42\u9a8c\u6536\u96c6&#xff1a;\u5148\u8981\u6c42\u751f\u6210\u4ee3\u7801\u80fd\u7f16\u8bd1\u3001\u80fd\u8fc7 testbench\u3002 \u53ef\u4ee5\u8fc1\u79fb\u5230 CRG\u3001\u4f4e\u529f\u8017\u3001reset\/clock domain \u5c0f\u6a21\u5757\u751f\u6210&#xff0c;\u4f46\u5fc5\u987b\u8865\u9f50\u9886\u57df testbench\u3002 \u5c40\u9650\u662f\u9898\u76ee\u76f8\u5bf9\u77ed\u5c0f&#xff0c;\u79bb\u771f\u5b9e repository\u3001IP \u4f9d\u8d56\u548c\u590d\u6742\u9a8c\u8bc1\u73af\u5883\u4ecd\u6709\u8ddd\u79bb\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>RTLFixer \u00b7 CraftRTL \u00b7 ScaleRTL \u00b7 CVDP \u00b7 RTLLM<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b Evaluation Framework\u3001pass&#064;k \u8bbe\u8ba1\u3001BLEU \u4e0e\u529f\u80fd\u6b63\u786e\u6027\u7684\u5dee\u5f02\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>ChipNeMo: Domain-Adapted LLMs for Chip Design<\/h4>\n<p>2023 \u00b7 ICML 2024 \/ arXiv \u00b7 AI4Design \u00b7 Domain LLM \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110947-6a71c87ba59ad.png\" alt=\"Paper 02 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.5&#xff5c;Figure 5: Chip Domain Benchmark Result for ChipNeMo.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110948-6a71c87c09f63.png\" alt=\"Paper 02 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.7&#xff5c;Table 2: EDA Script Generation Evaluation Benchmarks<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u901a\u7528\u5927\u6a21\u578b\u4e0d\u80fd\u76f4\u63a5\u7406\u89e3\u82af\u7247\u5de5\u7a0b\u8bed\u6599&#xff0c;\u4f4e\u6210\u672c\u9886\u57df\u9002\u914d\u80fd\u663e\u8457\u63d0\u5347\u5de5\u7a0b\u52a9\u624b\u3001EDA \u811a\u672c\u548c bug \u5206\u6790\u80fd\u529b\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u82af\u7247\u8bbe\u8ba1\u8bed\u6599\u5305\u542b\u79c1\u6709\u6587\u6863\u3001RTL\u3001EDA \u811a\u672c\u3001bug \u8bb0\u5f55\u3001\u5de5\u5177\u65e5\u5fd7\u548c\u5927\u91cf\u7f29\u5199\u3002\u901a\u7528 LLM \u5728\u81ea\u7136\u8bed\u8a00\u80fd\u529b\u4e0a\u5f3a&#xff0c;\u4f46\u5bf9\u8fd9\u4e9b\u8bed\u6599\u7684 tokenization\u3001\u68c0\u7d22 grounding \u548c\u5de5\u7a0b\u8bed\u5883\u4e0d\u591f\u7a33\u5b9a&#xff1b;\u5982\u679c\u76f4\u63a5\u63a5\u5165\u751f\u4ea7\u6d41\u7a0b&#xff0c;\u4f1a\u51fa\u73b0\u672f\u8bed\u8bef\u89e3\u3001API \u5e7b\u89c9\u548c\u56de\u7b54\u4e0d\u843d\u5730\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u4ee5 LLaMA2 \u4e3a\u57fa\u5ea7&#xff0c;\u52a0\u5165 domain-adaptive tokenizer&#xff0c;\u51cf\u5c11\u82af\u7247\u672f\u8bed\u548c\u4fe1\u53f7\u540d\u88ab\u8fc7\u5ea6\u5207\u788e\u3002<\/li>\n<li>\u4f7f\u7528\u7ea6 24B chip design tokens \u505a continued pretraining&#xff0c;\u5e76\u7528\u9886\u57df instruction data \u505a\u5bf9\u9f50\u3002<\/li>\n<li>\u9488\u5bf9 engineering assistant chatbot \u5f15\u5165\u9886\u57df retriever \u548c RAG&#xff0c;\u63d0\u5347\u56de\u7b54\u7684\u53ef\u8ffd\u6eaf\u6027\u3002<\/li>\n<li>\u9009\u62e9\u5de5\u7a0b\u52a9\u624b\u3001EDA script generation\u3001bug summarization\/analysis \u4e09\u7c7b\u5de5\u4e1a\u7528\u4f8b\u8bc4\u4f30\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u82af\u7247\u5de5\u7a0b\u77e5\u8bc6\u4e0d\u53ea\u662f\u672f\u8bed\u8868&#xff0c;\u800c\u662f\u6587\u6863\u3001\u4ee3\u7801\u3001\u811a\u672c\u3001bug\u3001\u5de5\u5177\u65e5\u5fd7\u7ec4\u6210\u7684\u79c1\u6709\u8bed\u5883\u3002<\/li>\n<li>\u9886\u57df\u9002\u914d\u7684\u6536\u76ca\u6765\u81ea tokenizer\u3001DAPT\u3001RAG\u3001instruction data \u7684\u7ec4\u5408&#xff0c;\u800c\u4e0d\u662f\u5355\u4e00\u5fae\u8c03\u3002<\/li>\n<li>EDA Copilot \u7684\u53ef\u7528\u6027\u53d6\u51b3\u4e8e grounding&#xff1a;\u56de\u7b54\u5fc5\u987b\u80fd\u56de\u5230\u5185\u90e8\u6587\u6863\u548c\u5de5\u5177\u4e8b\u5b9e\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u7528\u7ea6 24B chip design tokens \u505a continued pretraining\u3002<\/li>\n<li>\u5f15\u5165 domain-adaptive tokenizer&#xff0c;\u964d\u4f4e\u786c\u4ef6\u672f\u8bed\u548c\u4fe1\u53f7\u540d\u7684\u5207\u788e\u635f\u5931\u3002<\/li>\n<li>\u8986\u76d6\u5de5\u7a0b\u52a9\u624b\u3001EDA \u811a\u672c\u3001bug \u6458\u8981\u4e09\u7c7b\u5de5\u4e1a\u7528\u4f8b\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u7ed3\u8bba\u79f0 7B\/13B\/70B ChipNeMo \u76f8\u6bd4 LLaMA2 \u57fa\u5ea7\u6709\u660e\u663e\u63d0\u5347\u3002<\/li>\n<li>ChipNeMo-70B \u5728\u5de5\u7a0b\u52a9\u624b\u548c EDA \u811a\u672c\u751f\u6210\u4e24\u4e2a\u7528\u4f8b\u4e0a\u8d85\u8fc7 GPT-4\u3002<\/li>\n<li>\u53ea\u4f7f\u7528\u76f8\u5bf9\u5c11\u91cf\u989d\u5916\u9884\u8bad\u7ec3\u7b97\u529b&#xff0c;\u4f53\u73b0\u9886\u57df\u7ee7\u7eed\u9884\u8bad\u7ec3\u7684\u6295\u5165\u4ea7\u51fa\u6bd4\u3002<\/li>\n<li>\u9886\u57df\u7ee7\u7eed\u9884\u8bad\u7ec3\u53ef\u80fd\u524a\u5f31\u6307\u4ee4\u9075\u5faa&#xff0c;\u540e\u7eed ChipAlign \u9488\u5bf9\u8be5\u95ee\u9898\u8865\u6d1e\u3002<\/li>\n<li>\u5bf9\u4f01\u4e1a\u79c1\u6709\u8bed\u6599\u4f9d\u8d56\u5f3a&#xff0c;\u516c\u5f00\u590d\u73b0\u96be\u3002<\/li>\n<li>\u6a21\u578b\u77e5\u8bc6\u4ecd\u53ef\u80fd\u8fc7\u671f&#xff0c;\u5de5\u5177 API \u548c\u9879\u76ee flow \u9700\u8981\u5b9e\u65f6\u68c0\u7d22\u7ea6\u675f\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; ChipNeMo \u662f NVIDIA AI4Design \u8def\u7ebf\u91cc\u7684\u201c\u6a21\u578b\u5e95\u5ea7\u201d\u8bba\u6587\u3002\u5b83\u7684\u6838\u5fc3\u5224\u65ad\u662f&#xff1a;\u82af\u7247\u8bbe\u8ba1\u4e0d\u662f\u666e\u901a\u4ee3\u7801\u751f\u6210&#xff0c;\u9886\u57df\u8bed\u6599\u548c\u5de5\u5177\u8bed\u5883\u672c\u8eab\u5c31\u662f\u80fd\u529b\u6765\u6e90\u3002\u8bba\u6587\u4e5f\u8bf4\u660e\u4e00\u4e2a\u73b0\u5b9e\u4e8b\u5b9e&#xff1a;\u5f88\u591a\u4f01\u4e1a\u4e0d\u53ef\u80fd\u628a\u79c1\u6709 RTL \u548c bug \u6570\u636e\u5168\u90e8\u4ea4\u7ed9\u5916\u90e8\u95ed\u6e90\u6a21\u578b&#xff0c;\u56e0\u6b64\u5185\u90e8\u9886\u57df\u6a21\u578b\u3001\u9886\u57df\u68c0\u7d22\u548c\u5b89\u5168\u90e8\u7f72\u4f1a\u957f\u671f\u5b58\u5728\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u5bf9\u4f01\u4e1a\u843d\u5730\u6700\u6709\u542f\u53d1\u7684\u662f\u201c\u5c11\u91cf\u989d\u5916\u9884\u8bad\u7ec3\u7b97\u529b\u6362\u5927\u91cf\u9886\u57df\u6536\u76ca\u201d&#xff0c;\u4e0d\u662f\u76f2\u76ee\u4ece\u96f6\u8bad\u7ec3\u3002 EDA \u811a\u672c\u3001bug \u6458\u8981\u3001\u6587\u6863\u95ee\u7b54\u662f\u6bd4\u7aef\u5230\u7aef RTL \u751f\u6210\u66f4\u65e9\u53ef\u843d\u5730\u7684\u573a\u666f\u3002 \u5c40\u9650\u662f\u9886\u57df\u9002\u914d\u53ef\u80fd\u727a\u7272 instruction following&#xff0c;\u540e\u7eed ChipAlign \u6b63\u662f\u89e3\u51b3\u8fd9\u4e00\u526f\u4f5c\u7528\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>ChipAlign \u00b7 ScaleRTL \u00b7 ACE-RTL \u00b7 JARVIS<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b domain adaptation pipeline\u3001\u4e09\u4e2a\u5de5\u4e1a\u7528\u4f8b\u3001RAG \u4e0e retriever \u7684\u4f5c\u7528\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>RTLFixer: Automatically Fixing RTL Syntax Errors with Large Language Models<\/h4>\n<p>2023 \u00b7 DAC 2024 \/ arXiv \u00b7 AI4Design \u00b7 Tool Loop \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110948-6a71c87c5d149.png\" alt=\"Paper 03 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.5&#xff5c;Figure 4: VerilogEval pass&#064;1 results prior (inner) and post (outer) syntax error fixing with RTLFixer.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110948-6a71c87cbd3e3.png\" alt=\"Paper 03 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.6&#xff5c;Detected table region by pdfplumber fallback, page 6, table 1, rows 5.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>LLM \u751f\u6210 Verilog \u7684\u5927\u91cf\u5931\u8d25\u6765\u81ea\u8bed\u6cd5\u9519\u8bef&#xff0c;\u7f16\u8bd1\u5668\u53cd\u9988\u3001RAG \u548c ReAct \u80fd\u5f62\u6210\u81ea\u52a8\u4fee\u590d\u95ed\u73af\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u5355\u6b21\u63d0\u793a\u751f\u6210 RTL \u65f6&#xff0c;\u6a21\u578b\u5e38\u72af\u7aef\u53e3\u5bbd\u5ea6\u3001always \u5757\u3001\u8d4b\u503c\u7c7b\u578b\u3001endmodule\u3001\u58f0\u660e\u4f4d\u7f6e\u7b49 Verilog \u8bed\u6cd5\u9519\u8bef\u3002\u8bed\u6cd5\u9519\u8bef\u867d\u7136\u4f4e\u7ea7&#xff0c;\u5374\u4f1a\u76f4\u63a5\u963b\u65ad\u4eff\u771f\u548c\u540e\u7eed\u529f\u80fd\u9a8c\u8bc1\u3002\u9760\u4eba\u5de5\u628a\u7f16\u8bd1\u65e5\u5fd7\u590d\u5236\u7ed9\u6a21\u578b\u5f88\u6162&#xff0c;\u4e5f\u4e0d\u53ef\u89c4\u6a21\u5316\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u5148\u8ba9 LLM \u751f\u6210 Verilog&#xff0c;\u518d\u8c03\u7528\u7f16\u8bd1\u5668\u5f97\u5230\u5177\u4f53\u9519\u8bef\u65e5\u5fd7\u3002<\/li>\n<li>\u7528 RAG \u68c0\u7d22 Verilog \u8bed\u6cd5\u548c\u4fee\u590d\u89c4\u5219&#xff0c;\u628a\u5916\u90e8\u4e13\u5bb6\u77e5\u8bc6\u6ce8\u5165 prompt\u3002<\/li>\n<li>\u7528 ReAct \u5f0f observe-think-act \u5faa\u73af&#xff0c;\u8ba9\u6a21\u578b\u5206\u6b65\u5206\u6790\u9519\u8bef\u5e76\u63d0\u4ea4 patch\u3002<\/li>\n<li>\u5728 VerilogEval \u548c RTLLM \u76f8\u5173\u4efb\u52a1\u4e0a\u8bc4\u4f30\u8bed\u6cd5\u4fee\u590d\u6210\u529f\u7387\u53ca pass&#064;1 \u63d0\u5347\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>EDA \u5de5\u5177\u53cd\u9988\u5e94\u4f5c\u4e3a agent \u7684\u73af\u5883\u89c2\u6d4b&#xff0c;\u800c\u4e0d\u662f\u6700\u7ec8\u9a8c\u6536\u624d\u4f7f\u7528\u3002<\/li>\n<li>\u8bed\u6cd5\u9519\u8bef\u662f RTL \u751f\u6210\u7684\u4f4e\u5c42\u74f6\u9888&#xff0c;\u5148\u89e3\u51b3\u5b83\u624d\u80fd\u8fdb\u5165\u529f\u80fd\u8c03\u8bd5\u3002<\/li>\n<li>RAG \u7684\u4ef7\u503c\u4e0d\u662f\u589e\u52a0\u77e5\u8bc6\u91cf&#xff0c;\u800c\u662f\u628a\u4fee\u590d\u884c\u4e3a\u7ea6\u675f\u5230 Verilog \u89c4\u5219\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u628a compiler feedback\u3001RAG\u3001ReAct \u7ec4\u5408\u6210\u81ea\u52a8\u8c03\u8bd5\u5faa\u73af\u3002<\/li>\n<li>\u9762\u5411 Verilog \u8bed\u6cd5\u5931\u8d25\u6784\u5efa\u9519\u8bef\u4fee\u590d\u6570\u636e\u96c6\u3002<\/li>\n<li>\u8bc1\u660e\u5de5\u5177\u53cd\u9988\u95ed\u73af\u6bd4\u5355\u6b21 prompt \u66f4\u7a33\u5b9a\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a\u7ea6 55% LLM-generated Verilog \u9519\u8bef\u4e3a syntax-related\u3002<\/li>\n<li>RTLFixer \u62a5\u544a\u8bed\u6cd5\u9519\u8bef\u4fee\u590d\u6210\u529f\u7387\u9ad8\u8fbe 98.5%\u3002<\/li>\n<li>\u5728 VerilogEval-Machine \u548c VerilogEval-Human \u4e0a\u5206\u522b\u5e26\u6765 pass&#064;1 \u63d0\u5347\u3002<\/li>\n<li>\u4e3b\u8981\u89e3\u51b3 syntax&#xff0c;\u4e0d\u7b49\u540c\u4e8e\u529f\u80fd\u6b63\u786e\u3002<\/li>\n<li>\u9519\u8bef\u65e5\u5fd7\u8d28\u91cf\u51b3\u5b9a\u4fee\u590d\u8d28\u91cf&#xff1b;\u65e5\u5fd7\u8fc7\u957f\u4f1a\u6c61\u67d3\u4e0a\u4e0b\u6587\u3002<\/li>\n<li>\u5bf9\u5de5\u5177\u7248\u672c\u3001\u7f16\u8bd1\u547d\u4ee4\u548c include path \u6709\u4f9d\u8d56\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; RTLFixer \u662f\u4ece\u201c\u751f\u6210\u5668\u201d\u8d70\u5411\u201c\u8c03\u8bd5\u4ee3\u7406\u201d\u7684\u7b2c\u4e00\u6b65\u3002\u5b83\u8bc1\u660e EDA \u5de5\u5177\u672c\u8eab\u53ef\u4ee5\u6210\u4e3a\u6a21\u578b\u7684\u5916\u90e8\u611f\u77e5\u5668&#xff1a;\u7f16\u8bd1\u5668\u4e0d\u662f\u6700\u540e\u9a8c\u6536&#xff0c;\u800c\u662f\u4e2d\u95f4\u53cd\u9988\u6e90\u3002\u8fd9\u4e2a\u6a21\u5f0f\u540e\u6765\u5728 VerilogCoder \u7684\u4eff\u771f\/\u6ce2\u5f62\u8ffd\u8e2a\u3001JARVIS \u7684\u811a\u672c\u7f16\u8bd1\u5668\u3001Timing Agent \u7684 STA report \u91cc\u4e0d\u65ad\u6269\u5c55\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u4efb\u4f55\u5185\u90e8 RTL \u751f\u6210\u5de5\u5177\u7684\u6700\u4f4e\u6210\u672c\u589e\u5f3a&#xff1a;\u628a\u7f16\u8bd1\u9519\u8bef\u53d8\u6210\u7ed3\u6784\u5316\u53cd\u9988\u5faa\u73af\u3002 \u5bf9\u8bed\u6cd5\u7c7b\u95ee\u9898\u6536\u76ca\u9ad8&#xff0c;\u4f46\u5bf9\u529f\u80fd\u9519\u8bef\u3001\u534f\u8bae\u9519\u8bef\u3001\u65f6\u5e8f\/CDC\/RDC \u7c7b\u95ee\u9898\u80fd\u529b\u6709\u9650\u3002 \u9700\u8981\u63a7\u5236\u4fee\u590d\u8f6e\u6570\u548c\u65e5\u5fd7\u6ce8\u5165\u957f\u5ea6&#xff0c;\u5426\u5219\u5bb9\u6613\u8fdb\u5165\u65e0\u6548\u91cd\u8bd5\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>VerilogEval \u00b7 VerilogCoder \u00b7 JARVIS \u00b7 PRO-V-R1<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b ReAct loop\u3001RAG expert rules\u3001compiler feedback \u5982\u4f55\u5199\u5165 prompt\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>PyHDL-Eval: An LLM Evaluation Framework for Hardware Design Using Python-Embedded DSLs<\/h4>\n<p>2024 \u00b7 MLCAD 2024 \u00b7 AI4Design \u00b7 Benchmark \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110948-6a71c87ce7634.png\" alt=\"Paper 04 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.3&#xff5c;Figure 2: PyHDL-Eval Framework \u2013 ICL &#061; in-context learning; iverilog &#061; Icarus Verilog; PyDSL &#061; one of five Python-embedded DSLs: PyMTL3, PyRTL, MyHDL, Migen, Amaranth.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110949-6a71c87d586f9.png\" alt=\"Paper 04 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.5&#xff5c;Table 2: Average Pass Rate \u2013 Pass rate is averaged across 20 samples per problem across all 168 PyHDL-Eval problems.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u786c\u4ef6\u8bbe\u8ba1\u4e0d\u53ea\u5199 Verilog&#xff0c;Python-embedded HDL\/DSL \u4e5f\u9700\u8981\u72ec\u7acb\u8bc4\u6d4b&#xff0c;\u4e14 ICL \u5bf9\u8fd9\u4e9b\u4f4e\u9891 DSL \u7279\u522b\u91cd\u8981\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>PyMTL3\u3001PyRTL\u3001MyHDL\u3001Migen\u3001Amaranth \u7b49 Python HDL\/DSL \u80fd\u63d0\u5347\u786c\u4ef6\u8bbe\u8ba1\u751f\u4ea7\u7387&#xff0c;\u4f46\u4e92\u8054\u7f51\u4e0a\u76f8\u5173\u4ee3\u7801\u91cf\u8fdc\u5c11\u4e8e Verilog\u3002\u901a\u7528 LLM \u5bf9\u8fd9\u4e9b DSL \u7684\u8bed\u6cd5\u3001\u6784\u9020\u4e60\u60ef\u3001\u6d4b\u8bd5\u6d41\u7a0b\u90fd\u4e0d\u719f&#xff0c;\u5bb9\u6613\u628a Python \u8f6f\u4ef6\u4e60\u60ef\u8bef\u5957\u5230\u786c\u4ef6 DSL\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6784\u5efa 168 \u4e2a specification-to-RTL\/DSL \u95ee\u9898\u3002<\/li>\n<li>\u4e3a\u6bcf\u9898\u63d0\u4f9b Verilog reference\u3001Verilog testbench\u3001Python tests \u548c workflow orchestration scripts\u3002<\/li>\n<li>\u6bd4\u8f83 CodeGemma\u3001Llama3\u3001GPT-4\u3001GPT-4 Turbo \u7b49\u6a21\u578b\u5728 Verilog \u548c\u591a\u4e2a Python DSL \u4e0a\u7684\u8868\u73b0\u3002<\/li>\n<li>\u91cd\u70b9\u5206\u6790 in-context learning \u5bf9\u5c0f\u6a21\u578b\u548c\u4f4e\u9891 DSL \u7684\u5e2e\u52a9\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>AI4Design \u7684\u5bf9\u8c61\u4e0d\u5e94\u5c40\u9650\u4e8e Verilog&#xff0c;DSL\/generator \u540c\u6837\u662f\u786c\u4ef6\u8868\u8fbe\u3002<\/li>\n<li>\u4f4e\u9891 DSL \u7684\u4e3b\u8981\u74f6\u9888\u662f\u8bed\u6599\u7a00\u7f3a&#xff0c;\u56e0\u6b64\u793a\u4f8b\u4e0a\u4e0b\u6587\u548c\u6587\u6863 grounding \u7279\u522b\u5173\u952e\u3002<\/li>\n<li>\u540c\u4e00\u89c4\u683c\u8de8 HDL\/DSL \u7684\u8bc4\u6d4b\u80fd\u63ed\u793a\u6a21\u578b\u662f\u5426\u771f\u6b63\u7406\u89e3\u786c\u4ef6\u610f\u56fe\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u8986\u76d6 PyMTL3\u3001PyRTL\u3001MyHDL\u3001Migen\u3001Amaranth \u7b49 Python-embedded DSL\u3002<\/li>\n<li>\u4e3a\u6bcf\u9898\u914d\u5957 Verilog reference\u3001testbench\u3001Python tests \u548c workflow scripts\u3002<\/li>\n<li>\u7cfb\u7edf\u6bd4\u8f83 ICL \u5bf9 Verilog \u4e0e Python DSL \u7684\u5dee\u5f02\u5316\u6536\u76ca\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a CodeGemma 7B \u5728 Verilog \u4e0a\u901a\u8fc7 ICL \u6709\u660e\u663e\u63d0\u5347\u3002<\/li>\n<li>Llama3 70B \u5728 PyMTL3 \u4e0a\u4e5f\u901a\u8fc7 ICL \u4ece\u63a5\u8fd1\u4e0d\u53ef\u7528\u63d0\u5347\u5230\u53ef\u6bd4\u8f83\u6c34\u5e73\u3002<\/li>\n<li>\u7ed3\u8bba\u6307\u51fa\u6a21\u578b\u666e\u904d\u66f4\u64c5\u957f Verilog&#xff0c;Python HDL\/DSL \u5dee\u8ddd\u4ecd\u660e\u663e\u3002<\/li>\n<li>\u4efb\u52a1\u4ecd\u4ee5\u5c0f\u8bbe\u8ba1\u4e3a\u4e3b&#xff0c;\u5c1a\u672a\u8986\u76d6\u5927\u578b\u53c2\u6570\u5316 generator \u5de5\u7a0b\u3002<\/li>\n<li>Python DSL \u7684\u6b63\u786e\u6027\u65e2\u53d7 Python \u8bed\u4e49\u5f71\u54cd&#xff0c;\u4e5f\u53d7\u786c\u4ef6 elaboration \u8bed\u4e49\u5f71\u54cd\u3002<\/li>\n<li>DSL \u6587\u6863\u7248\u672c\u53d8\u5316\u4f1a\u5f71\u54cd\u6a21\u578b\u8f93\u51fa\u6709\u6548\u6027\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; \u8fd9\u7bc7\u8bba\u6587\u6269\u5c55\u4e86 VerilogEval \u7684\u8fb9\u754c\u3002\u5b83\u63d0\u9192\u6211\u4eec&#xff1a;AI4Design \u4e0d\u5e94\u628a\u201c\u786c\u4ef6\u4ee3\u7801\u201d\u7b49\u540c\u4e8e Verilog\u3002\u672a\u6765\u82af\u7247\u8bbe\u8ba1\u53ef\u80fd\u8d8a\u6765\u8d8a\u591a\u7528 Python-based generators\u3001DSL\u3001\u914d\u7f6e\u811a\u672c\u548c\u53c2\u6570\u5316\u751f\u6210\u5668\u8868\u8fbe&#xff0c;\u56e0\u6b64\u8bc4\u6d4b\u4e5f\u5fc5\u987b\u8986\u76d6\u8fd9\u4e9b\u975e\u4f20\u7edf RTL \u8868\u793a\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u5bf9 Chisel\u3001SpinalHDL\u3001PyMTL\u3001\u5185\u90e8 Python generator \u7b49\u56e2\u961f\u6709\u76f4\u63a5\u53c2\u8003\u4ef7\u503c\u3002 \u4f4e\u9891 DSL \u66f4\u4f9d\u8d56\u793a\u4f8b\u3001\u6587\u6863\u68c0\u7d22\u548c\u6a21\u677f\u5316\u4e0a\u4e0b\u6587&#xff0c;\u4e0d\u80fd\u53ea\u9760\u6a21\u578b\u53c2\u6570\u8bb0\u5fc6\u3002 \u5c40\u9650\u662f\u4efb\u52a1\u4ecd\u504f\u5c0f\u6a21\u5757&#xff0c;\u5c1a\u672a\u5145\u5206\u8986\u76d6\u5927\u578b generator repo \u7684\u4f9d\u8d56\u548c\u6784\u5efa\u95ee\u9898\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>VerilogEval \u00b7 RTLLM \u00b7 CVDP<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b\u8de8 DSL benchmark \u7ec4\u7ec7\u65b9\u5f0f\u3001ICL \u793a\u4f8b\u8bbe\u8ba1\u548c workflow orchestration\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Revisiting VerilogEval: Newer LLMs, In-Context Learning, and Specification-to-RTL Tasks<\/h4>\n<p>2024 \u00b7 TODAES 2025 \/ arXiv \u00b7 AI4Design \u00b7 Benchmark \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110949-6a71c87dc2a44.png\" alt=\"Paper 05 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.5&#xff5c;Fig. 2. Overview of VerilogEval v2 flow.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110950-6a71c87e24771.png\" alt=\"Paper 05 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.10&#xff5c;Table 1. Types of failures supported by automatic failure classification.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u6a21\u578b\u5feb\u901f\u8fdb\u6b65\u540e&#xff0c;\u539f VerilogEval \u9700\u8981\u5347\u7ea7\u5230\u66f4\u771f\u5b9e\u7684 specification-to-RTL \u548c instruction-tuned \u573a\u666f\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u4e00\u5e74\u5185\u65b0\u6a21\u578b\u80fd\u529b\u5927\u5e45\u63d0\u5347&#xff0c;\u65e7 benchmark \u4f1a\u51fa\u73b0\u533a\u5206\u5ea6\u4e0b\u964d\u3001prompt \u5f62\u5f0f\u4e0d\u9002\u914d instruction-tuned \u6a21\u578b\u3001\u4efb\u52a1\u5f62\u6001\u504f code completion \u7684\u95ee\u9898\u3002\u5982\u679c\u8bc4\u6d4b\u4e0d\u5347\u7ea7&#xff0c;\u7814\u7a76\u4f1a\u88ab\u8fc7\u65f6\u6307\u6807\u8bef\u5bfc\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6bd4\u8f83 GPT-4o\u3001GPT-4 Turbo\u3001Llama3.1\u3001Mistral Large\u3001DeepSeek Coder\u3001CodeGemma\u3001RTL-Coder \u7b49\u65b0\u6a21\u578b\u3002<\/li>\n<li>\u5c06\u4efb\u52a1\u4ece code completion \u6269\u5c55\u5230 specification-to-RTL\u3002<\/li>\n<li>\u5206\u6790 zero-shot \u4e0e in-context learning \u5728\u4e0d\u540c\u6a21\u578b\u548c\u4efb\u52a1\u4e0a\u7684\u6536\u76ca\u5dee\u5f02\u3002<\/li>\n<li>\u7ee7\u7eed\u4fdd\u7559\u53ef\u6267\u884c\u4eff\u771f\u8bc4\u5206&#xff0c;\u907f\u514d\u8bc4\u6d4b\u9000\u56de\u6587\u672c\u76f8\u4f3c\u5ea6\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>benchmark \u5fc5\u987b\u968f\u6a21\u578b\u80fd\u529b\u6f14\u8fdb&#xff0c;\u5426\u5219\u4f1a\u4ece\u533a\u5206\u5668\u53d8\u6210\u8363\u8a89\u699c\u3002<\/li>\n<li>spec-to-RTL \u6bd4 code completion \u66f4\u63a5\u8fd1\u771f\u5b9e\u8bbe\u8ba1\u5165\u53e3\u3002<\/li>\n<li>ICL \u662f\u4efb\u52a1\/\u6a21\u578b\u76f8\u5173\u7b56\u7565&#xff0c;\u800c\u4e0d\u662f\u7edf\u4e00\u589e\u76ca\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u91cd\u65b0\u8bc4\u6d4b GPT-4o\u3001Llama3.1\u3001Mistral\u3001DeepSeek Coder\u3001RTL-Coder \u7b49\u6a21\u578b\u3002<\/li>\n<li>\u628a VerilogEval \u6269\u5c55\u5230 instruction\/specification-driven \u4efb\u52a1\u3002<\/li>\n<li>\u4fdd\u7559\u4eff\u771f\u53ef\u6267\u884c\u8bc4\u5206&#xff0c;\u907f\u514d\u6307\u6807\u6f02\u79fb\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a GPT-4o \u5728 spec-to-RTL \u4e0a\u8fbe\u5230\u7ea6 63% pass rate\u3002<\/li>\n<li>Llama3.1 405B \u63a5\u8fd1 GPT-4o&#xff0c;\u8bf4\u660e\u5f00\u6e90 frontier model \u5df2\u63a5\u8fd1\u5546\u7528\u6a21\u578b\u3002<\/li>\n<li>\u5c0f\u578b domain-specific RTL-Coder \u5728\u53c2\u6570\u6548\u7387\u4e0a\u6709\u7ade\u4e89\u529b\u3002<\/li>\n<li>\u65b0\u6a21\u578b\u63d0\u5347\u4e0d\u4ee3\u8868\u771f\u5b9e\u5de5\u7a0b\u81ea\u52a8\u5316\u5df2\u7ecf\u89e3\u51b3\u3002<\/li>\n<li>prompt \u6a21\u677f\u3001ICL \u793a\u4f8b\u548c\u8bc4\u6d4b harness \u4f1a\u663e\u8457\u5f71\u54cd\u7ed3\u679c\u3002<\/li>\n<li>\u4ecd\u7f3a\u5c11\u591a\u6587\u4ef6\u3001\u5de5\u5177\u94fe\u548c\u5de5\u7a0b\u4f9d\u8d56\u538b\u529b\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; \u8fd9\u7bc7\u8bba\u6587\u5728 NVIDIA \u8def\u7ebf\u91cc\u627f\u62c5\u201c\u6821\u51c6\u5c3a\u66f4\u65b0\u201d\u7684\u89d2\u8272\u3002\u5b83\u8bf4\u660e benchmark \u4e0d\u662f\u4e00\u6b21\u6027\u8d44\u4ea7&#xff0c;\u800c\u8981\u968f\u6a21\u578b\u3001prompt\u3001\u4efb\u52a1\u5f62\u6001\u540c\u6b65\u6f14\u8fdb\u3002\u5c24\u5176\u91cd\u8981\u7684\u662f&#xff0c;\u5b83\u628a spec-to-RTL \u653e\u5230\u4e2d\u5fc3\u4f4d\u7f6e&#xff0c;\u8fd9\u6bd4\u8865\u5168\u534a\u6bb5\u4ee3\u7801\u66f4\u8d34\u8fd1\u5de5\u7a0b\u5e08\u771f\u5b9e\u9700\u6c42\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u5185\u90e8\u8bc4\u6d4b\u96c6\u5fc5\u987b\u5b9a\u671f\u5237\u65b0&#xff0c;\u5426\u5219\u4f1a\u9ad8\u4f30\u6a21\u578b\u5728\u771f\u5b9e\u8bbe\u8ba1\u4efb\u52a1\u4e2d\u7684\u80fd\u529b\u3002 ICL \u4e0d\u662f\u7a33\u5b9a\u4e07\u80fd\u836f&#xff0c;\u5e94\u6309\u6a21\u578b\u548c\u4efb\u52a1\u5206\u522b\u8bc4\u4f30\u6210\u672c\u6536\u76ca\u3002 \u4ecd\u9700\u8fdb\u4e00\u6b65\u6269\u5927\u5230 repo-level\u3001\u4f9d\u8d56\u5b9a\u4f4d\u548c\u591a\u6587\u4ef6\u7f16\u8f91&#xff0c;\u540e\u7eed CVDP \u627f\u63a5\u4e86\u8fd9\u4e00\u95ee\u9898\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>VerilogEval \u00b7 CVDP \u00b7 ScaleRTL \u00b7 ACE-RTL<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b v1 \u5230 v2 \u7684\u4efb\u52a1\u53d8\u5316&#xff0c;\u4ee5\u53ca\u4e0d\u540c\u6a21\u578b\u5728 ICL \u4e0b\u7684\u589e\u76ca\u5dee\u5f02\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>CraftRTL: High-quality Synthetic Data Generation for Verilog Code Models<\/h4>\n<p>2024 \u00b7 ICLR 2025 \/ arXiv \u00b7 AI4Design \u00b7 Data Engine \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110950-6a71c87e792c2.png\" alt=\"Paper 06 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.5&#xff5c;Figure 2: State transition logic.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110950-6a71c87ed0d32.png\" alt=\"Paper 06 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.3&#xff5c;Table 2: pass&#064;1 results on VerilogEval sam- pled with temperature of 0.8.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>RTL \u8bad\u7ec3\u6570\u636e\u4e0d\u80fd\u53ea\u8ffd\u6c42\u89c4\u6a21&#xff0c;\u5fc5\u987b correct-by-construction&#xff0c;\u5e76\u8986\u76d6 K-map\u3001FSM\u3001waveform \u7b49\u975e\u6587\u672c\u89c4\u683c\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u5f00\u6e90 Verilog \u6570\u636e\u5c11\u3001\u8d28\u91cf\u53c2\u5dee&#xff0c;\u5408\u6210\u6570\u636e\u53c8\u5bb9\u6613\u628a\u9519\u8bef\u6a21\u5f0f\u653e\u5927\u3002\u6a21\u578b\u5728 Karnaugh map\u3001\u72b6\u6001\u8f6c\u79fb\u56fe\u3001\u6ce2\u5f62\u7b49\u975e\u7eaf\u6587\u672c\u89c4\u683c\u4e0a\u5c24\u5176\u5bb9\u6613\u5931\u8d25&#xff0c;\u5e76\u4e14\u8bad\u7ec3 checkpoint \u4f1a\u968f\u673a\u4ea7\u751f\u5c0f\u9519\u8bef&#xff0c;\u5bfc\u81f4\u5fae\u8c03\u4e0d\u7a33\u5b9a\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u7528\u7a0b\u5e8f\u5316\u751f\u6210\u5668\u6784\u9020 correct-by-construction \u6570\u636e&#xff0c;\u786e\u4fdd\u89c4\u683c\u548c\u4ee3\u7801\u5929\u7136\u4e00\u81f4\u3002<\/li>\n<li>\u8986\u76d6 Karnaugh map\u3001FSM\u3001waveform \u7b49\u975e\u6587\u672c\u8868\u793a&#xff0c;\u8865\u9f50\u666e\u901a\u6587\u672c prompt \u7684\u76f2\u533a\u3002<\/li>\n<li>\u6536\u96c6\u4e0d\u540c checkpoint \u7684\u9519\u8bef\u62a5\u544a&#xff0c;\u5e76\u5c06\u9519\u8bef\u6ce8\u5165\u5f00\u6e90\u4ee3\u7801\u751f\u6210 targeted code repair \u6570\u636e\u3002<\/li>\n<li>\u5fae\u8c03 StarCoder2-15B&#xff0c;\u5e76\u5728 VerilogEval \u4e0e RTLLM \u4e0a\u6bd4\u8f83\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u786c\u4ef6\u8bad\u7ec3\u6570\u636e\u5e94\u662f\u53ef\u751f\u6210\u3001\u53ef\u8bc1\u660e\u3001\u53ef\u4fee\u590d\u7684\u5de5\u7a0b\u5bf9\u8c61\u3002<\/li>\n<li>\u975e\u6587\u672c\u89c4\u683c\u662f RTL \u751f\u6210\u7684\u91cd\u8981\u771f\u5b9e\u5165\u53e3&#xff0c;\u4e0d\u80fd\u88ab\u7eaf\u6587\u672c benchmark \u8986\u76d6\u3002<\/li>\n<li>\u9519\u8bef\u4fee\u590d\u6570\u636e\u5e94\u6765\u81ea\u6a21\u578b\u771f\u5b9e\u5931\u8d25\u6a21\u5f0f&#xff0c;\u800c\u4e0d\u662f\u4eba\u5de5\u60f3\u8c61\u9519\u8bef\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u7a0b\u5e8f\u5316\u751f\u6210 K-map\u3001FSM\u3001waveform \u7b49 correct-by-construction \u6570\u636e\u3002<\/li>\n<li>\u5206\u6790 fine-tuned model \u7684\u5931\u8d25\u6a21\u5f0f\u540e\u6784\u9020 targeted repair data\u3002<\/li>\n<li>\u7528 StarCoder2-15B \u5c55\u793a\u9ad8\u8d28\u91cf\u5408\u6210\u6570\u636e\u5bf9 RTL \u751f\u6210\u7684\u63d0\u5347\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a\u5728 VerilogEval-Machine\u3001VerilogEval-Human\u3001RTLLM \u4e0a\u5747\u8d85\u8fc7\u524d\u5e8f SOTA\u3002<\/li>\n<li>\u9519\u8bef\u6ce8\u5165\u6765\u81ea checkpoint error reports&#xff0c;\u4f7f\u4fee\u590d\u6837\u672c\u66f4\u8d34\u8fd1\u771f\u5b9e\u6a21\u578b\u5931\u8d25\u3002<\/li>\n<li>\u5f3a\u8c03\u6570\u636e\u8d28\u91cf\u4e0e\u8986\u76d6\u7c7b\u578b\u6bd4\u5355\u7eaf\u6269\u5927\u8bed\u6599\u66f4\u5173\u952e\u3002<\/li>\n<li>\u7a0b\u5e8f\u5316\u6570\u636e\u751f\u6210\u5668\u672c\u8eab\u9700\u8981\u786c\u4ef6\u4e13\u5bb6\u7ef4\u62a4\u3002<\/li>\n<li>correct-by-construction \u4e0d\u4fdd\u8bc1\u8986\u76d6\u771f\u5b9e\u5de5\u4e1a\u4ee3\u7801\u98ce\u683c\u3002<\/li>\n<li>\u975e\u6587\u672c\u89c4\u683c\u8d8a\u590d\u6742&#xff0c;\u81ea\u52a8\u751f\u6210\u548c\u9a8c\u8bc1\u6210\u672c\u8d8a\u9ad8\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; CraftRTL \u662f\u6570\u636e\u5de5\u7a0b\u8bba\u6587&#xff0c;\u4f46\u5b83\u7684\u601d\u60f3\u975e\u5e38\u5de5\u7a0b\u5316&#xff1a;\u4e0d\u8981\u6307\u671b LLM \u81ea\u5df1\u751f\u6210\u9ad8\u8d28\u91cf\u8bad\u7ec3\u96c6&#xff0c;\u786c\u4ef6\u6570\u636e\u5fc5\u987b\u5e26\u5f62\u5f0f\u7ea6\u675f\u548c\u53ef\u6267\u884c\u6821\u9a8c\u3002\u5bf9 EDA \u6765\u8bf4&#xff0c;\u6570\u636e\u6784\u9020\u672c\u8eab\u5c31\u662f\u9a8c\u8bc1\u5de5\u7a0b\u7684\u4e00\u90e8\u5206\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u7528\u6765\u6307\u5bfc\u5185\u90e8 CRG\u3001reset\u3001protocol block \u7684 synthetic dataset \u6784\u9020\u3002 \u5173\u952e\u662f\u628a\u9886\u57df\u89c4\u5219\u5199\u6210\u751f\u6210\u5668\u548c checker&#xff0c;\u800c\u4e0d\u662f\u53ea\u5199\u81ea\u7136\u8bed\u8a00\u9898\u76ee\u3002 \u5c40\u9650\u662f correct-by-construction \u6570\u636e\u4ecd\u53ef\u80fd\u8986\u76d6\u4e0d\u5230\u771f\u5b9e IP \u7684\u590d\u6742\u4e0a\u4e0b\u6587\u548c\u98ce\u683c\u5dee\u5f02\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>VerilogEval \u00b7 ScaleRTL \u00b7 CVDP \u00b7 ACE-RTL<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b\u6570\u636e\u751f\u6210\u5668\u3001\u9519\u8bef\u6ce8\u5165\u7b56\u7565\u548c targeted code repair \u6570\u636e\u6784\u9020\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and AST-based Waveform Tracing Tool<\/h4>\n<p>2024 \u00b7 AAAI 2025 \/ arXiv \u00b7 AI4Design \u00b7 Agent \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110951-6a71c87f28d54.png\" alt=\"Paper 07 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.4&#xff5c;Figure 3: An illustration of task-driven circuit relation graph retrieval agent reasoning and interacting with the developed TCRG retrieval tool to enrich the task with the relevant circuit and signal descriptions.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110951-6a71c87f81409.png\" alt=\"Paper 07 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.6&#xff5c;Table 1: Pass-rates of recent large language models (i.e., non-agentic method) and the proposed VerilogCoder. We run the VerilogCoder once for each problem in the bench- mark. The pass-rates of VerilogCoder (agentic method) &#061; #passed case\/#total case. For the<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>RTL agent \u9700\u8981\u4efb\u52a1\u56fe\u89c4\u5212\u548c\u6ce2\u5f62\/AST \u7ea7\u8c03\u8bd5\u80fd\u529b&#xff0c;\u4e0d\u80fd\u53ea\u9760\u81ea\u7136\u8bed\u8a00\u53cd\u601d\u4fee\u529f\u80fd\u9519\u8bef\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>RTLFixer \u80fd\u4fee\u8bed\u6cd5&#xff0c;\u4f46\u529f\u80fd\u9519\u8bef\u66f4\u96be\u3002\u529f\u80fd\u9519\u8bef\u5f80\u5f80\u4f53\u73b0\u5728\u7279\u5b9a\u7aef\u53e3\u3001\u72b6\u6001\u673a\u8def\u5f84\u3001\u7ec4\u5408\u6761\u4ef6\u6216\u65f6\u5e8f\u884c\u4e3a\u4e0a&#xff0c;\u7f16\u8bd1\u5668\u65e5\u5fd7\u4e0d\u80fd\u544a\u8bc9\u6a21\u578b\u54ea\u91cc\u9519\u3002\u6ca1\u6709\u4fe1\u53f7\u7ea7\u5b9a\u4f4d\u80fd\u529b&#xff0c;agent \u53ea\u80fd\u76f2\u76ee\u91cd\u5199\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u63d0\u51fa Task and Circuit Relation Graph&#xff0c;\u7528\u4efb\u52a1\/\u7535\u8def\u5173\u7cfb\u56fe\u8f85\u52a9\u751f\u6210\u6574\u4f53\u8ba1\u5212\u3002<\/li>\n<li>\u6784\u5efa\u591a agent \u6d41\u7a0b&#xff0c;\u63a5\u5165 syntax checker\u3001simulator\u3001waveform tracer\u3002<\/li>\n<li>\u8bbe\u8ba1 AST-based waveform tracing&#xff0c;\u4ece\u5931\u8d25\u8f93\u51fa\u6cbf\u8bed\u6cd5\u6811\u548c\u4fe1\u53f7\u4f9d\u8d56\u53cd\u5411\u5b9a\u4f4d\u76f8\u5173\u903b\u8f91\u3002<\/li>\n<li>\u5c06\u5b9a\u4f4d\u7ed3\u679c\u6ce8\u5165\u4fee\u590d prompt&#xff0c;\u5f62\u6210\u8bed\u6cd5\u548c\u529f\u80fd\u53cc\u95ed\u73af\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u529f\u80fd\u8c03\u8bd5\u9700\u8981\u4fe1\u53f7\u7ea7\u56e0\u679c\u5b9a\u4f4d&#xff0c;\u800c\u4e0d\u662f\u81ea\u7136\u8bed\u8a00\u81ea\u6211\u53cd\u601d\u3002<\/li>\n<li>\u4efb\u52a1\u56fe\u89c4\u5212\u80fd\u628a\u6a21\u5757\u63cf\u8ff0\u8f6c\u5316\u4e3a\u53ef\u68c0\u67e5\u7684\u5c40\u90e8\u5b50\u76ee\u6807\u3002<\/li>\n<li>\u6ce2\u5f62\u8ffd\u8e2a\u662f RTL agent \u4ece syntax-level \u8fdb\u5165 behavior-level \u7684\u5173\u952e\u611f\u77e5\u80fd\u529b\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u63d0\u51fa Task and Circuit Relation Graph \u505a\u5168\u5c40\u4efb\u52a1\u89c4\u5212\u3002<\/li>\n<li>\u5f15\u5165 AST-based waveform tracing \u5b9a\u4f4d\u529f\u80fd\u9519\u8bef\u76f8\u5173\u4fe1\u53f7\u3002<\/li>\n<li>\u628a syntax checker\u3001simulator\u3001waveform tracer \u4e32\u5165\u591a agent \u6d41\u7a0b\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u6307\u51fa\u6b64\u524d\u4ec5\u9760 simulator\/RAG \u4fee syntax \u96be\u4ee5\u63d0\u5347 functional success\u3002<\/li>\n<li>AST-WT \u80fd\u4ece\u5931\u8d25\u8f93\u51fa\u56de\u6eaf\u5230\u76f8\u5173 RTL \u903b\u8f91&#xff0c;\u51cf\u5c11\u76f2\u76ee\u91cd\u5199\u3002<\/li>\n<li>\u591a\u5de5\u5177\u534f\u540c\u663e\u8457\u63d0\u5347\u590d\u6742 Verilog \u4efb\u52a1\u901a\u8fc7\u7387\u3002<\/li>\n<li>\u9700\u8981 testbench \u66b4\u9732\u5931\u8d25\u884c\u4e3a&#xff0c;\u5426\u5219 tracing \u65e0\u4ece\u5f00\u59cb\u3002<\/li>\n<li>\u590d\u6742\u5c42\u6b21\u548c generated logic \u4f1a\u589e\u52a0 AST\/\u4fe1\u53f7\u6620\u5c04\u96be\u5ea6\u3002<\/li>\n<li>\u771f\u5b9e\u9879\u76ee\u9700\u8981 waveform database\u3001\u5c42\u6b21\u8def\u5f84\u548c\u4eff\u771f\u6027\u80fd\u5de5\u7a0b\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; VerilogCoder \u662f NVIDIA \u4ece\u201c\u7f16\u8bd1\u95ed\u73af\u201d\u8d70\u5411\u201c\u4eff\u771f\u95ed\u73af\u201d\u7684\u5173\u952e\u8282\u70b9\u3002\u5b83\u8bf4\u660e RTL agent \u7684\u91cd\u8981\u80fd\u529b\u4e0d\u662f\u4f1a\u8bf4\u201c\u8ba9\u6211\u68c0\u67e5\u4e00\u4e0b\u201d&#xff0c;\u800c\u662f\u80fd\u628a\u5931\u8d25\u6ce2\u5f62\u5207\u7247\u3001\u4fe1\u53f7\u4f9d\u8d56\u548c AST \u7ed3\u6784\u8f6c\u5316\u6210\u53ef\u4fee\u6539\u7684\u4ee3\u7801\u533a\u57df\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u5bf9\u5185\u90e8 RTL debug agent \u5f88\u5173\u952e&#xff1a;\u5fc5\u987b\u505a waveform slicing \u548c signal dependency&#xff0c;\u800c\u4e0d\u662f\u53ea\u603b\u7ed3\u65e5\u5fd7\u3002 \u53ef\u8fc1\u79fb\u5230 reset release\u3001clock gating enable\u3001handshake FSM \u7b49\u529f\u80fd\u8c03\u8bd5\u573a\u666f\u3002 \u5c40\u9650\u662f\u6ce2\u5f62\u5b9a\u4f4d\u4f9d\u8d56 testbench \u8986\u76d6\u548c\u4fe1\u53f7\u53ef\u89c2\u6d4b\u6027&#xff0c;\u5bf9\u590d\u6742 SoC \u8fd8\u9700\u8981\u5c42\u6b21\u5316\u62bd\u8c61\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>RTLFixer \u00b7 ACE-RTL \u00b7 Trace2Skill<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b TCRG \u548c AST-WT&#xff0c;\u4e24\u8005\u5206\u522b\u89e3\u51b3\u89c4\u5212\u4e0e\u8bca\u65ad\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>FVEval: Understanding Language Model Capabilities in Formal Verification of Digital Hardware<\/h4>\n<p>2024 \u00b7 DATE 2025 \/ arXiv \u00b7 AI4EDA \u00b7 Verification \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110952-6a71c880033f5.png\" alt=\"Paper 08 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.25&#xff5c;Figure 12: Details of the prompt given to LLMs for the NL2SVA-Human benchmark.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110952-6a71c88055b9f.png\" alt=\"Paper 08 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.21&#xff5c;Table 6: Statistics describing the NL2SVA-Human benchmark. From industrial formal testbenches, we extract a total of 79 test cases of NL specification to SVA assertion.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>LLM \u505a\u786c\u4ef6\u9a8c\u8bc1\u4e0d\u80fd\u53ea\u770b assertion \u662f\u5426\u50cf\u6837&#xff0c;\u5fc5\u987b\u7528 FV \u5de5\u5177\u548c\u903b\u8f91\u7b49\u4ef7\u6027\u68c0\u67e5\u8bed\u4e49\u6b63\u786e\u6027\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u5f62\u5f0f\u9a8c\u8bc1\u5bf9\u82af\u7247\u8d28\u91cf\u5f88\u5173\u952e&#xff0c;\u4f46\u4e13\u5bb6\u6210\u672c\u9ad8\u3002LLM \u4f3c\u4e4e\u80fd\u5199 SVA&#xff0c;\u4f46\u81ea\u7136\u8bed\u8a00\u89c4\u683c\u5e38\u4e0d\u5b8c\u6574&#xff0c;RTL \u4fe1\u53f7\u5173\u7cfb\u590d\u6742&#xff0c;\u65ad\u8a00\u8bed\u6cd5\u6b63\u786e\u4e0d\u7b49\u4e8e\u9a8c\u8bc1\u8bed\u4e49\u6b63\u786e\u3002\u7f3a\u5c11 holistic evaluation \u4f1a\u8ba9\u7814\u7a76\u9ad8\u4f30\u6a21\u578b\u80fd\u529b\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u5c06 FV \u80fd\u529b\u62c6\u4e3a\u591a\u7c7b\u4efb\u52a1&#xff0c;\u5305\u62ec\u81ea\u7136\u8bed\u8a00\u5230 SVA\u3001RTL \u7406\u89e3\u548c\u9a8c\u8bc1\u63a8\u7406\u3002<\/li>\n<li>\u4f7f\u7528\u4e13\u5bb6 collateral \u4e0e\u53ef\u6269\u5c55 synthetic examples \u6784\u9020\u6d4b\u8bd5\u5b9e\u4f8b\u3002<\/li>\n<li>\u63a5\u5165\u5de5\u4e1a\u7ea7 formal verification tool \u505a\u7aef\u5230\u7aef\u81ea\u52a8\u8bc4\u5206\u3002<\/li>\n<li>\u63d0\u51fa model-generated assertion \u4e0e ground-truth assertion \u7684\u903b\u8f91\u7b49\u4ef7\u68c0\u67e5\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u5f62\u5f0f\u9a8c\u8bc1\u4efb\u52a1\u5fc5\u987b\u8bc4\u4ef7\u903b\u8f91\u8bed\u4e49&#xff0c;\u4e0d\u5e94\u8bc4\u4ef7 assertion \u6587\u672c\u5916\u89c2\u3002<\/li>\n<li>FV assistant \u7684\u771f\u6b63\u96be\u70b9\u662f\u4ece\u4e0d\u5b8c\u6574\u89c4\u683c\u6062\u590d\u9a8c\u8bc1\u610f\u56fe\u3002<\/li>\n<li>\u5de5\u4e1a\u7ea7 formal tool \u662f LLM \u9a8c\u8bc1\u80fd\u529b\u8bc4\u4f30\u7684\u5fc5\u8981 oracle\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u63d0\u51fa\u9996\u4e2a\u9762\u5411\u786c\u4ef6\u5f62\u5f0f\u9a8c\u8bc1\u7684\u7efc\u5408 LLM benchmark\u3002<\/li>\n<li>\u8986\u76d6 SVA \u751f\u6210\u3001RTL \u7406\u89e3\u3001\u9a8c\u8bc1\u63a8\u7406\u7b49\u591a\u5b50\u4efb\u52a1\u3002<\/li>\n<li>\u5f15\u5165 assertion equivalence checking \u8861\u91cf\u8bed\u4e49\u6b63\u786e\u6027\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u57fa\u4e8e\u4e13\u5bb6 collateral \u548c synthetic examples \u6784\u5efa\u6d4b\u8bd5\u96c6\u3002<\/li>\n<li>\u4f7f\u7528 industry-standard formal verification tool \u505a\u81ea\u52a8\u8bc4\u4f30\u3002<\/li>\n<li>\u7ed3\u8bba\u663e\u793a\u5f53\u524d LLM \u5728 FV \u4e0a\u4ecd\u6709\u663e\u8457\u77ed\u677f\u3002<\/li>\n<li>\u516c\u5f00 benchmark \u96be\u8986\u76d6\u4f01\u4e1a\u771f\u5b9e FV collateral\u3002<\/li>\n<li>\u7b49\u4ef7\u68c0\u67e5\u4f9d\u8d56 ground-truth assertion \u8d28\u91cf\u3002<\/li>\n<li>\u751f\u6210 assertion \u4e0d\u7b49\u540c\u4e8e\u5b8c\u6210\u9a8c\u8bc1\u8ba1\u5212\u548c\u8986\u76d6\u6536\u655b\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; FVEval \u628a AI4EDA \u4ece\u8bbe\u8ba1\u751f\u6210\u62c9\u5230\u9a8c\u8bc1\u8bed\u4e49\u5c42\u3002\u5b83\u7684\u91cd\u8981\u6027\u5728\u4e8e\u5f3a\u8c03\u201cassertion \u7684\u6587\u5b57\u5f62\u5f0f\u201d\u6ca1\u6709\u610f\u4e49&#xff0c;\u771f\u6b63\u8981\u770b\u662f\u5426\u8bc1\u660e\u4e86\u6b63\u786e\u7684\u6027\u8d28\u3001\u662f\u5426\u6f0f\u6389 corner case\u3001\u662f\u5426\u4e0e RTL \u884c\u4e3a\u7b49\u4ef7\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u5185\u90e8\u9a8c\u8bc1\u52a9\u624b\u5e94\u4f18\u5148\u8f93\u51fa\u53ef\u88ab formal tool \u68c0\u67e5\u7684 SVA\/cover\/property&#xff0c;\u800c\u4e0d\u662f\u53ea\u7ed9\u5efa\u8bae\u3002 \u9002\u5408\u5efa\u7acb property generation\u3001assertion review\u3001FV collateral QA \u7684\u8bc4\u6d4b\u57fa\u7ebf\u3002 \u5c40\u9650\u662f FV \u5de5\u5177\u548c\u5de5\u4e1a design collateral \u4f9d\u7136\u79c1\u6709\u6602\u8d35&#xff0c;\u516c\u5f00\u590d\u73b0\u96be\u5ea6\u8f83\u9ad8\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>AssertionForge \u00b7 PRO-V-R1 \u00b7 CVDP<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b\u4e09\u4e2a FV \u5b50\u4efb\u52a1\u3001\u903b\u8f91\u7b49\u4ef7\u6307\u6807\u548c formal tool \u8bc4\u4ef7\u6d41\u7a0b\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>ChipAlign: Instruction Alignment in Large Language Models for Chip Design via Geodesic Interpolation<\/h4>\n<p>2024 \u00b7 DAC 2025 \/ arXiv \u00b7 AI4Design \u00b7 Domain LLM \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110952-6a71c880b022e.png\" alt=\"Paper 09 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.3&#xff5c;Figure 3: An overview of ChipAlign.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110953-6a71c881093b3.png\" alt=\"Paper 09 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.5&#xff5c;Table 1: ROUGE-L scores on the OpenROAD QA benchmark \u2014 \u2217denotes the results sourced from [16].<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u82af\u7247\u9886\u57df\u6a21\u578b\u53ef\u80fd\u61c2\u9886\u57df\u4f46\u4e0d\u542c\u6307\u4ee4&#xff0c;training-free geodesic model merging \u80fd\u540c\u65f6\u4fdd\u7559\u9886\u57df\u77e5\u8bc6\u548c\u6307\u4ee4\u9075\u5faa\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>ChipNeMo \u7c7b\u6a21\u578b\u901a\u8fc7\u9886\u57df\u7ee7\u7eed\u9884\u8bad\u7ec3\u589e\u5f3a\u4e86\u82af\u7247\u77e5\u8bc6&#xff0c;\u4f46\u53ef\u80fd\u524a\u5f31 instruction following\u3002\u5de5\u7a0b\u52a9\u624b\u5982\u679c\u4e0d\u4e25\u683c\u9075\u5faa\u7528\u6237\u7ea6\u675f&#xff0c;\u4f1a\u5728\u811a\u672c\u751f\u6210\u3001\u95ee\u7b54\u548c\u8c03\u8bd5\u4efb\u52a1\u4e2d\u4ea7\u751f\u4e0d\u53ef\u63a7\u8f93\u51fa\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u5c06 chip-specific LLM \u4e0e general instruction-aligned LLM \u8fdb\u884c\u6743\u91cd\u878d\u5408\u3002<\/li>\n<li>\u4e0d\u7528\u989d\u5916\u8bad\u7ec3&#xff0c;\u800c\u662f\u7528 geodesic interpolation \/ Slerp \u5728\u6743\u91cd\u7a7a\u95f4\u6cbf\u6d4b\u5730\u7ebf\u5408\u5e76\u3002<\/li>\n<li>\u7528 IFEval \u8861\u91cf\u6307\u4ee4\u9075\u5faa&#xff0c;\u7528 OpenROAD QA \u548c production-level chip QA \u8861\u91cf\u9886\u57df\u80fd\u529b\u3002<\/li>\n<li>\u6bd4\u8f83\u7ebf\u6027\u5408\u5e76\u3001\u51e0\u4f55\u5408\u5e76\u548c\u539f\u59cb ChipNeMo \u7b49 baseline\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u9886\u57df\u77e5\u8bc6\u548c\u6307\u4ee4\u9075\u5faa\u662f\u4e24\u4e2a\u4e0d\u540c\u80fd\u529b\u8f74&#xff0c;\u5355\u72ec\u5f3a\u5316\u4e00\u4e2a\u4f1a\u4f24\u5bb3\u5de5\u7a0b\u53ef\u7528\u6027\u3002<\/li>\n<li>\u6a21\u578b\u5408\u5e76\u53ef\u4ee5\u4f5c\u4e3a\u4f4e\u6210\u672c\u540e\u8bad\u7ec3\u66ff\u4ee3&#xff0c;\u4f46\u5fc5\u987b\u5c0a\u91cd\u6743\u91cd\u7a7a\u95f4\u51e0\u4f55\u3002<\/li>\n<li>EDA assistant \u7684\u6838\u5fc3\u4e0d\u662f\u4f1a\u56de\u7b54&#xff0c;\u800c\u662f\u6309\u7ea6\u675f\u683c\u5f0f\u53ef\u9760\u6267\u884c\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u7528 training-free model merging \u878d\u5408 chip-specific \u4e0e instruction-aligned LLM\u3002<\/li>\n<li>\u91c7\u7528 geodesic interpolation \/ Slerp&#xff0c;\u800c\u4e0d\u662f\u7b80\u5355\u7ebf\u6027\u63d2\u503c\u3002<\/li>\n<li>\u540c\u65f6\u8bc4\u4f30 IFEval\u3001OpenROAD QA \u548c production-level chip QA\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a IFEval \u6700\u9ad8\u63d0\u5347\u7ea6 26.6%\u3002<\/li>\n<li>OpenROAD QA \u548c\u751f\u4ea7\u7ea7 chip QA \u4e5f\u6709 instruction-involved gains\u3002<\/li>\n<li>\u76f8\u6bd4 ChipNeMo&#xff0c;ChipAlign \u66f4\u597d\u5e73\u8861\u9886\u57df\u80fd\u529b\u548c\u542c\u6307\u4ee4\u80fd\u529b\u3002<\/li>\n<li>\u6a21\u578b\u5408\u5e76\u4e0d\u80fd\u4fee\u590d\u6240\u6709 hallucination \u548c tool API \u9519\u8bef\u3002<\/li>\n<li>\u4e0d\u540c\u57fa\u5ea7\u6a21\u578b\u7684\u67b6\u6784\/\u8bad\u7ec3\u5dee\u5f02\u4f1a\u5f71\u54cd\u5408\u5e76\u6548\u679c\u3002<\/li>\n<li>\u5b89\u5168\u3001\u6743\u9650\u548c\u8f93\u51fa\u7ea6\u675f\u4ecd\u9700\u5916\u90e8 guardrail\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; ChipAlign \u56de\u7b54\u4e86\u9886\u57df\u6a21\u578b\u843d\u5730\u65f6\u7684\u5173\u952e\u77db\u76fe&#xff1a;\u77e5\u8bc6\u548c\u53ef\u63a7\u6027\u90fd\u91cd\u8981\u3002\u53ea\u61c2\u82af\u7247\u4f46\u4e0d\u9075\u5b88\u683c\u5f0f\u3001\u8fb9\u754c\u548c\u5de5\u5177\u7ea6\u675f&#xff0c;\u4e0d\u80fd\u4f5c\u4e3a\u5de5\u7a0b\u52a9\u624b&#xff1b;\u53ea\u542c\u8bdd\u4f46\u4e0d\u61c2\u82af\u7247&#xff0c;\u53c8\u65e0\u6cd5\u89e3\u51b3\u771f\u5b9e EDA \u95ee\u9898\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u4f01\u4e1a\u5185\u90e8\u628a\u901a\u7528 instruction model \u4e0e\u79c1\u6709\u9886\u57df\u6a21\u578b\u878d\u5408&#xff0c;\u964d\u4f4e\u91cd\u65b0 RLHF\/SFT \u6210\u672c\u3002 \u5bf9 EDA Copilot \u7279\u522b\u91cd\u8981&#xff0c;\u56e0\u4e3a\u811a\u672c\u548c\u9a8c\u8bc1\u4efb\u52a1\u901a\u5e38\u6709\u4e25\u683c\u8f93\u51fa\u683c\u5f0f\u3002 \u5c40\u9650\u662f model merging \u4e0d\u662f\u4e07\u80fd&#xff0c;\u6df1\u5c42\u80fd\u529b\u51b2\u7a81\u548c\u5b89\u5168\u8fb9\u754c\u4ecd\u9700\u8bc4\u6d4b\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>ChipNeMo \u00b7 OpenROAD QA \u00b7 JARVIS<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b geodesic interpolation\u3001IFEval \u548c chip QA \u4e09\u8005\u5982\u4f55\u5171\u540c\u8861\u91cf\u80fd\u529b\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>AssertionForge: Enhancing Formal Verification Assertion Generation with Structured Representation of Specifications and RTL<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4EDA \u00b7 Verification \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110953-6a71c88163011.png\" alt=\"Paper 10 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.6&#xff5c;Figure 2 visualizes KGs constructed from the OPEN- MSP430 design specification, contrasting the impact of our domain-specific schema.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110953-6a71c881c199d.png\" alt=\"Paper 10 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.3&#xff5c;Detected table region by pdfplumber fallback, page 3, table 2, rows 4.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>SVA \u751f\u6210\u9700\u8981\u540c\u65f6\u7406\u89e3\u89c4\u683c\u548c RTL&#xff0c;\u5b9e\u73b0\u77e5\u8bc6\u56fe\u8c31\u6bd4\u5355\u7eaf\u89c4\u683c\u6587\u672c\u66f4\u80fd\u6355\u6349\u591a\u4fe1\u53f7\u4ea4\u4e92\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u81ea\u7136\u8bed\u8a00\u89c4\u683c\u901a\u5e38\u7f3a\u5c11\u5185\u90e8\u4fe1\u53f7\u3001\u72b6\u6001\u6761\u4ef6\u548c\u5b9e\u73b0\u7ec6\u8282\u3002\u53ea\u4ece spec \u751f\u6210 assertion \u5bb9\u6613\u6f0f\u6389 RTL \u4e2d\u771f\u5b9e\u5b58\u5728\u7684\u4f9d\u8d56\u5173\u7cfb&#xff0c;\u5bfc\u81f4\u65ad\u8a00\u770b\u4f3c\u5408\u7406\u4f46\u8986\u76d6\u4e0d\u5230\u5173\u952e\u884c\u4e3a\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u4ece specification \u548c RTL \u540c\u65f6\u6784\u5efa\u786c\u4ef6\u4e13\u7528 Knowledge Graph\u3002<\/li>\n<li>\u5b9a\u4e49\u6a21\u5757\u3001\u4fe1\u53f7\u3001\u72b6\u6001\u3001\u6761\u4ef6\u3001\u6570\u636e\u4f9d\u8d56\u7b49\u5b9e\u4f53\u548c\u5173\u7cfb\u3002<\/li>\n<li>\u7528 global summary\u3001signal-specific retriever\u3001guided random walk with adaptive sampling \u5408\u6210\u591a\u5206\u8fa8\u7387\u4e0a\u4e0b\u6587\u3002<\/li>\n<li>\u751f\u6210 SystemVerilog Assertions&#xff0c;\u5e76\u7528 formal tool \u8bc4\u4f30 proven assertions\u3001coverage \u548c\u8d28\u91cf\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>SVA \u751f\u6210\u5fc5\u987b\u540c\u65f6\u5efa\u6a21\u8bbe\u8ba1\u610f\u56fe\u548c\u5b9e\u73b0\u7ec6\u8282\u3002<\/li>\n<li>\u77e5\u8bc6\u56fe\u8c31\u662f\u628a spec \u4e0e RTL \u5bf9\u9f50\u7684\u4e00\u79cd\u53ef\u89e3\u91ca\u4e2d\u95f4\u8868\u793a\u3002<\/li>\n<li>\u591a\u5206\u8fa8\u7387\u4e0a\u4e0b\u6587\u6bd4\u628a\u5b8c\u6574 spec\/RTL \u585e\u8fdb prompt \u66f4\u53ef\u9760\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u4ece specification \u548c RTL \u6784\u5efa\u7edf\u4e00\u786c\u4ef6 Knowledge Graph\u3002<\/li>\n<li>\u5b9a\u4e49\u786c\u4ef6\u5b9e\u4f53\u3001\u4fe1\u53f7\u5173\u7cfb\u3001\u72b6\u6001\u6761\u4ef6\u548c\u6570\u636e\u4f9d\u8d56\u3002<\/li>\n<li>\u7528 global summary\u3001signal retriever\u3001guided random walk \u5408\u6210\u9a8c\u8bc1\u4e0a\u4e0b\u6587\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u6307\u51fa\u53ea\u770b spec \u7684 ASSERTLLM \u7c7b\u65b9\u6cd5\u4f1a\u6f0f\u6389 RTL \u5185\u90e8\u4fe1\u53f7\u4ea4\u4e92\u3002<\/li>\n<li>KG \u8def\u5f84\u80fd\u5e2e\u52a9\u53d1\u73b0\u591a\u4fe1\u53f7\u5173\u8054 property\u3002<\/li>\n<li>formal tool \u8bc4\u4f30 proven assertions\u3001coverage \u548c assertion quality\u3002<\/li>\n<li>KG \u62bd\u53d6\u9519\u8bef\u4f1a\u76f4\u63a5\u8bef\u5bfc assertion \u751f\u6210\u3002<\/li>\n<li>\u590d\u6742 SystemVerilog \u7279\u6027\u3001\u5b8f\u548c generate \u7ed3\u6784\u9700\u8981\u66f4\u5f3a parser\u3002<\/li>\n<li>coverage \u63d0\u5347\u4e0d\u5fc5\u7136\u7b49\u4e8e\u9a8c\u8bc1\u8ba1\u5212\u5b8c\u6574\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; AssertionForge \u5ef6\u7eed FVEval&#xff0c;\u4f46\u4ece\u201c\u8bc4\u6d4b\u201d\u8d70\u5411\u201c\u751f\u6210\u65b9\u6cd5\u201d\u3002\u5b83\u7684\u5173\u952e\u4e0d\u662f\u8ba9 prompt \u66f4\u957f&#xff0c;\u800c\u662f\u628a\u89c4\u683c\u610f\u56fe\u548c RTL \u5b9e\u73b0\u90fd\u7ed3\u6784\u5316&#xff0c;\u518d\u628a\u76f8\u5173\u8def\u5f84\u5582\u7ed9\u6a21\u578b\u3002\u8fd9\u4e0e\u5de5\u7a0b\u5e08\u5199 assertion \u7684\u8ba4\u77e5\u8fc7\u7a0b\u4e00\u81f4&#xff1a;\u5148\u7406\u89e3\u6574\u4f53\u529f\u80fd&#xff0c;\u518d\u627e\u76f8\u5173\u4fe1\u53f7&#xff0c;\u518d\u5199 property\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u5185\u90e8\u505a assertion assistant&#xff1a;\u8f93\u5165 spec &#043; RTL&#xff0c;\u8f93\u51fa\u5019\u9009 SVA \u548c\u5bf9\u5e94\u4fe1\u53f7\u8def\u5f84\u4f9d\u636e\u3002 \u77e5\u8bc6\u56fe\u8c31\u4e5f\u53ef\u590d\u7528\u5230 RTL \u7406\u89e3\u3001\u4ee3\u7801 review\u3001design intent recovery\u3002 \u5c40\u9650\u662f KG \u62bd\u53d6\u8d28\u91cf\u51b3\u5b9a\u4e0a\u9650&#xff1b;\u590d\u6742 RTL \u7684\u5c42\u6b21\u3001generate\u3001macro\u3001interface \u4ecd\u4f1a\u589e\u52a0\u96be\u5ea6\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>FVEval \u00b7 PRO-V-R1 \u00b7 Knowledge Graph<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b KG schema\u3001\u591a\u5206\u8fa8\u7387\u4e0a\u4e0b\u6587\u5408\u6210\u548c GRW-AS\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Marco: Configurable Graph-Based Task Solving and Multi-AI Agents Framework for Hardware Design<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4Design \u00b7 Agent Framework \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04julfx2aiuhl.png\" alt=\"Paper 11 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.2&#xff5c;Fig. 1: Graph-Based Task Solving illustration in Marco: Configurable Graph-Based Task Solving and Multi-AI Agents Framework.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04xb3oslck3hn.png\" alt=\"Paper 11 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.3&#xff5c;Detected table region by pdfplumber fallback, page 3, table 1, rows 9.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u786c\u4ef6 agent \u4e0d\u5e94\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u5355\u72ec\u62fc\u7cfb\u7edf&#xff0c;\u800c\u5e94\u6709\u7edf\u4e00\u7684\u56fe\u5f0f\u4efb\u52a1\u6c42\u89e3\u548c\u591a agent \u7f16\u6392\u6846\u67b6\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u524d\u9762\u7684 RTLFixer\u3001VerilogCoder\u3001Timing Agent\u3001JARVIS \u90fd\u662f\u4efb\u52a1\u4e13\u7528 agent\u3002\u771f\u5b9e EDA \u6d41\u7a0b\u8de8 synthesis\u3001verification\u3001physical design\u3001timing\u3001DRC\/LVS \u7b49\u591a\u4e2a\u73af\u8282&#xff0c;\u5982\u679c\u6bcf\u4e2a\u73af\u8282\u5355\u72ec\u505a demo&#xff0c;\u5f88\u96be\u7ec4\u6210\u751f\u4ea7\u7cfb\u7edf\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u63d0\u51fa configurable graph-based task solving&#xff0c;\u628a\u786c\u4ef6\u4efb\u52a1\u62c6\u6210\u56fe\u4e0a\u7684\u5b50\u4efb\u52a1\u3002<\/li>\n<li>\u6bcf\u4e2a\u5b50\u4efb\u52a1\u53ef\u7ed1\u5b9a\u4e0d\u540c agent\u3001\u5de5\u5177\u3001\u77e5\u8bc6\u6e90\u548c\u9a8c\u8bc1\u5668\u3002<\/li>\n<li>\u652f\u6301\u5355 agent\u3001\u591a agent\u3001\u591a\u6a21\u6001\u8f93\u5165\u548c\u9886\u57df\u5de5\u5177\u8c03\u7528\u3002<\/li>\n<li>\u5728 cell layout optimization\u3001Verilog syntax fixing\u3001Verilog\/DRC code generation\u3001timing debugging \u7b49\u4efb\u52a1\u4e0a\u793a\u8303\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>EDA agent \u5e73\u53f0\u5e94\u4ee5\u4efb\u52a1\u56fe\u4e3a\u63a7\u5236\u5e73\u9762&#xff0c;\u4ee5\u5de5\u5177\/verifier \u4e3a\u6267\u884c\u5e73\u9762\u3002<\/li>\n<li>\u591a agent \u7684\u4ef7\u503c\u5728\u4e8e\u4e13\u4e1a\u5206\u5de5\u548c\u53ef\u66ff\u6362\u5de5\u5177\u63a5\u53e3&#xff0c;\u800c\u4e0d\u53ea\u662f\u591a\u4e2a\u804a\u5929\u89d2\u8272\u3002<\/li>\n<li>\u6846\u67b6\u5316\u80fd\u529b\u51b3\u5b9a\u5355\u70b9 demo \u80fd\u5426\u53d8\u6210\u53ef\u7ef4\u62a4\u7cfb\u7edf\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u63d0\u51fa configurable graph-based task solving\u3002<\/li>\n<li>\u7edf\u4e00 layout\u3001Verilog\u3001DRC\u3001timing \u7b49\u591a\u7c7b\u786c\u4ef6\u4efb\u52a1\u3002<\/li>\n<li>\u5c55\u793a cell layout optimization\u3001Verilog\/DRC code generation\u3001timing analysis \u7b49\u6848\u4f8b\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a sequential cells \u4e0a\u9762\u79ef\u548c LVS\/DRC clean rate \u6709\u6536\u76ca\u3002<\/li>\n<li>\u6574\u5408 RTLFixer\u3001VerilogCoder \u7b49\u4efb\u52a1\u578b agent \u6210\u4e3a\u7edf\u4e00\u6846\u67b6\u3002<\/li>\n<li>\u7ed3\u8bba\u5f3a\u8c03\u672a\u6765\u8981\u8bad\u7ec3\u9ad8\u8d28\u91cf\u786c\u4ef6\u6570\u636e\u3001\u96c6\u6210 PPA loop \u548c self-learning memory\u3002<\/li>\n<li>\u6846\u67b6\u8bba\u6587\u7bc7\u5e45\u77ed&#xff0c;\u751f\u4ea7\u7ea7\u8c03\u5ea6\u3001\u5b89\u5168\u548c\u6210\u672c\u63a7\u5236\u7ec6\u8282\u4e0d\u8db3\u3002<\/li>\n<li>\u4efb\u52a1\u56fe\u8d28\u91cf\u9ad8\u5ea6\u4f9d\u8d56\u9886\u57df\u5efa\u6a21\u3002<\/li>\n<li>\u5de5\u5177 adapter \u548c verifier \u4e0d\u5b8c\u5584\u65f6&#xff0c;\u591a agent \u53ea\u4f1a\u653e\u5927\u566a\u58f0\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; Marco \u66f4\u50cf\u5e73\u53f0\u8bba\u6587\u3002\u5b83\u628a NVIDIA \u591a\u6761 agent \u7ebf\u7d22\u62bd\u8c61\u6210\u53ef\u7ec4\u5408\u67b6\u6784&#xff1a;\u4efb\u52a1\u56fe\u662f\u63a7\u5236\u5e73\u9762&#xff0c;\u9886\u57df\u5de5\u5177\u662f\u6267\u884c\u5e73\u9762&#xff0c;LLM \u662f\u63a8\u7406\u548c\u80f6\u6c34\u3002\u5bf9\u4f01\u4e1a\u6765\u8bf4&#xff0c;\u8fd9\u6bd4\u5355\u70b9 benchmark \u66f4\u63a5\u8fd1 AI4EDA \u5e73\u53f0\u5f62\u6001\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u5185\u90e8\u5e73\u53f0\u5e94\u6309 task graph\u3001tool adapter\u3001verifier\u3001memory\u3001trace store \u5212\u5206\u6a21\u5757\u3002 \u9002\u5408\u4f5c\u4e3a\u591a\u4e2a EDA agent \u7684\u8c03\u5ea6\u5c42&#xff0c;\u800c\u4e0d\u662f\u66ff\u4ee3\u6240\u6709\u5e95\u5c42\u5de5\u5177\u3002 \u5c40\u9650\u662f\u8bba\u6587\u7bc7\u5e45\u8f83\u77ed&#xff0c;\u771f\u5b9e\u5927\u89c4\u6a21\u90e8\u7f72\u4ecd\u9700\u8981\u6743\u9650\u3001\u5b89\u5168\u3001\u5e76\u53d1\u3001\u6210\u672c\u548c\u5ba1\u8ba1\u673a\u5236\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>VerilogCoder \u00b7 JARVIS \u00b7 Timing Agent \u00b7 Trace2Skill<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b task graph \u62bd\u8c61&#xff0c;\u4ee5\u53ca\u4e0d\u540c EDA task \u5982\u4f55\u88ab\u7edf\u4e00\u6302\u63a5\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Timing Analysis Agent: Autonomous MCMM Timing Debugging with Timing Debug Relation Graph<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4EDA \u00b7 Agent \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04pb22r4i21nr.png\" alt=\"Paper 12 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.3&#xff5c;Fig. 3: Examples of max and xtalk max timing report for a specific corner and mode.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-043ryvezhk0yk.png\" alt=\"Paper 12 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.6&#xff5c;Detected table region by pdfplumber fallback, page 6, table 1, rows 12.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>MCMM timing debug \u9700\u8981\u5c42\u6b21\u5316\u89c4\u5212\u3001\u62a5\u544a\u68c0\u7d22\u548c Timing Debug Relation Graph&#xff0c;\u800c\u4e0d\u662f\u628a\u6240\u6709 STA \u62a5\u544a\u585e\u8fdb\u4e0a\u4e0b\u6587\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u5148\u8fdb\u5de5\u827a\u4e0b MCMM timing report \u53c8\u957f\u53c8\u590d\u6742&#xff0c;\u5de5\u7a0b\u5e08\u9700\u8981\u8de8 corner\u3001mode\u3001path\u3001variation report\u3001constraint \u548c\u7ecf\u9a8c\u89c4\u5219\u5224\u65ad root cause\u3002\u901a\u7528 RAG \u5f88\u5bb9\u6613\u68c0\u7d22\u566a\u58f0&#xff0c;\u6216\u8005\u53ea\u603b\u7ed3\u62a5\u544a\u8868\u9762\u4fe1\u606f\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6784\u5efa Timing Debug Relation Graph&#xff0c;\u84b8\u998f timing debug trace \u548c\u4e13\u5bb6\u77e5\u8bc6\u3002<\/li>\n<li>\u8bbe\u8ba1 MCMM planner agent\u3001TDRG traversal agent\u3001expert report agent \u5206\u5c42\u6c42\u89e3\u3002<\/li>\n<li>\u63d0\u51fa Agentic RAG&#xff0c;\u5229\u7528 LLM \u7684\u4ee3\u7801\u80fd\u529b\u4ece\u62a5\u544a\u4e2d\u68c0\u7d22\u5fc5\u8981 timing \u4fe1\u606f\u5e76\u8fc7\u6ee4\u566a\u58f0\u3002<\/li>\n<li>\u5728 single-report \u548c multi-report benchmark \u4e0a\u8bc4\u4f30\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>Timing debug \u7684\u672c\u8d28\u662f\u8de8 report \u7684\u56e0\u679c\u56fe\u63a8\u7406&#xff0c;\u4e0d\u662f\u6587\u672c\u6458\u8981\u3002<\/li>\n<li>MCMM \u573a\u666f\u9700\u8981\u5148\u89c4\u5212 mode\/corner\/path&#xff0c;\u518d\u505a\u8bc1\u636e\u68c0\u7d22\u3002<\/li>\n<li>Agentic RAG \u5e94\u628a retrieval \u53d8\u6210\u53ef\u6267\u884c\u4fe1\u606f\u62bd\u53d6&#xff0c;\u800c\u4e0d\u662f\u76f8\u4f3c\u6bb5\u843d\u53ec\u56de\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u63d0\u51fa Timing Debug Relation Graph \u84b8\u998f timing debug trace\u3002<\/li>\n<li>\u6784\u5efa MCMM planner\u3001TDRG traversal\u3001expert report \u591a agent \u6d41\u7a0b\u3002<\/li>\n<li>\u5229\u7528 LLM coding ability \u4ece\u5546\u4e1a\u5de5\u5177\u62a5\u544a\u4e2d\u63d0\u53d6\u5fc5\u8981 timing \u4fe1\u606f\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a single-report benchmark \u5e73\u5747\u7ea6 98% pass rate\u3002<\/li>\n<li>multi-report benchmark \u7ea6 90% pass rate\u3002<\/li>\n<li>\u76f8\u6bd4\u5176\u4ed6 RAG \u6280\u672f\u9ad8\u51fa 46% \u4ee5\u4e0a\u3002<\/li>\n<li>\u4f9d\u8d56 STA report \u683c\u5f0f\u3001\u9879\u76ee timing methodology \u548c\u4e13\u5bb6\u77e5\u8bc6\u3002<\/li>\n<li>\u4e0d\u80fd\u66ff\u4ee3\u6700\u7ec8 signoff&#xff0c;\u53ea\u80fd\u8f85\u52a9\u5b9a\u4f4d\u548c\u89e3\u91ca\u3002<\/li>\n<li>\u5546\u4e1a\u5de5\u5177\u8f93\u51fa\u548c\u5185\u90e8\u8bbe\u8ba1\u6570\u636e\u9650\u5236\u516c\u5f00\u590d\u73b0\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; \u8fd9\u7bc7\u8bba\u6587\u628a agent \u5e26\u5230\u771f\u6b63\u540e\u7aef signoff \u573a\u666f\u3002\u5b83\u7684\u8981\u70b9\u662f&#xff1a;EDA report \u4e0d\u662f\u666e\u901a\u6587\u6863&#xff0c;\u5fc5\u987b\u7406\u89e3\u62a5\u544a\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u548c timing debug \u7684\u5de5\u7a0b\u6d41\u7a0b\u3002\u5173\u7cfb\u56fe\u8ba9 agent \u80fd\u50cf\u5de5\u7a0b\u5e08\u4e00\u6837\u5148\u95ee\u201c\u8be5\u770b\u54ea\u4e2a mode\/corner\/path\u201d&#xff0c;\u800c\u4e0d\u662f\u76f4\u63a5\u6458\u8981\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u5185\u90e8 STA debug assistant&#xff0c;\u4f46\u5fc5\u987b\u63a5\u5165\u771f\u5b9e report parser \u548c\u8def\u5f84\u7ea7\u6570\u636e\u7ed3\u6784\u3002 \u5bf9 CRG\/clock tree\/timing exception review \u6709\u8fc1\u79fb\u4ef7\u503c&#xff0c;\u56e0\u4e3a\u90fd\u9700\u8981\u8de8\u62a5\u544a\u5173\u7cfb\u63a8\u7406\u3002 \u5c40\u9650\u662f\u5546\u7528 STA \u5de5\u5177\u683c\u5f0f\u3001\u4fdd\u5bc6\u6570\u636e\u548c signoff \u8d23\u4efb\u8fb9\u754c\u4f1a\u5f71\u54cd\u843d\u5730\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>Marco \u00b7 JARVIS \u00b7 Synopsys.ai<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b TDRG \u5982\u4f55\u7ec4\u7ec7 report \u4e0e\u4e13\u5bb6 debug knowledge\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>JARVIS: A Multi-Agent Code Assistant for High-Quality EDA Script Generation<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4Design \u00b7 Agent \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04fxxmp3riyus.png\" alt=\"Paper 13 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.4&#xff5c;Fig. 3: Overview of the multi-agent flow, listing the vari- ous tools utilized to improve code quality: Code Generator, RuleEnforce, Code Compiler, Code Fixing Agent, RAG, and a Guardrail agent.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-042vndh2wjo1l.png\" alt=\"Paper 13 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.4&#xff5c;Detected table region by pdfplumber fallback, page 4, table 3, rows 14.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>EDA \u811a\u672c\u751f\u6210\u7684\u6838\u5fc3\u98ce\u9669\u662f API \u5e7b\u89c9\u548c\u89c4\u5219\u9519\u8bef&#xff0c;\u5fc5\u987b\u7528 custom compiler\u3001RAG\u3001\u4fee\u590d\u5de5\u5177\u548c\u591a agent \u53cd\u9988\u95ed\u73af\u7ea6\u675f\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>EDA \u811a\u672c\u901a\u5e38\u4f9d\u8d56\u4e13\u6709 API\u3001\u547d\u4ee4\u8bed\u6cd5\u3001\u8bbe\u8ba1\u89c4\u5219\u548c\u4e0a\u4e0b\u6587\u5bf9\u8c61\u3002LLM \u5f88\u5bb9\u6613\u751f\u6210\u770b\u4f3c\u5408\u7406\u4f46\u5de5\u5177\u4e0d\u63a5\u53d7\u7684\u811a\u672c&#xff0c;\u6216\u8005\u8c03\u7528\u4e0d\u5b58\u5728\u7684 API\u3002\u7eaf\u81ea\u7136\u8bed\u8a00 prompt \u96be\u4ee5\u63a7\u5236\u8fd9\u4e9b\u786c\u7ea6\u675f\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u8bad\u7ec3\/\u9002\u914d domain-specific LLM&#xff0c;\u5e76\u7528 synthetic data generation \u8865\u5145\u9886\u57df\u77e5\u8bc6\u3002<\/li>\n<li>\u6784\u5efa custom compiler \u505a\u7ed3\u6784\u9a8c\u8bc1\u3001\u89c4\u5219\u68c0\u67e5\u548c API \u6821\u9a8c\u3002<\/li>\n<li>\u63a5\u5165 advanced retrieval&#xff0c;\u4ece\u5de5\u5177\u624b\u518c\u548c\u89c4\u5219\u5e93\u83b7\u53d6 grounding\u3002<\/li>\n<li>\u4f7f\u7528 multi-episode\u3001multi-agent ReAct \u5faa\u73af&#xff0c;\u6839\u636e\u7f16\u8bd1\u53cd\u9988\u4e0d\u65ad\u4fee\u590d\u811a\u672c\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>EDA \u811a\u672c\u751f\u6210\u7684\u6b63\u786e\u6027\u5e94\u7531\u5de5\u5177 API schema \u548c compiler guardrail \u5b9a\u4e49\u3002<\/li>\n<li>LLM \u4e0d\u5e94\u9760\u8bb0\u5fc6\u4e13\u6709\u547d\u4ee4&#xff0c;\u800c\u5e94\u901a\u8fc7\u68c0\u7d22\u548c\u7f16\u8bd1\u53cd\u9988\u63a5\u5165\u5de5\u5177\u4e8b\u5b9e\u3002<\/li>\n<li>\u811a\u672c agent \u662f\u77ed\u671f\u6700\u5bb9\u6613\u4ea7\u4e1a\u5316\u7684 AI4Design \u573a\u666f\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u7ed3\u5408 domain-specific LLM\u3001synthetic data\u3001custom compiler\u3001retrieval \u548c code fixing\u3002<\/li>\n<li>\u4f7f\u7528 multi-episode multi-agent ReAct \u4f18\u5316\u811a\u672c\u3002<\/li>\n<li>\u9762\u5411 specialized EDA script generation \u89e3\u51b3 data scarcity \u4e0e hallucination\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a\u5728\u591a\u4e2a benchmark \u4e0a\u8d85\u8fc7\u5df2\u6709 domain-specific model\u3002<\/li>\n<li>custom compiler \u7528\u4e8e\u7ed3\u6784\u9a8c\u8bc1\u3001\u89c4\u5219\u68c0\u67e5\u548c API \u6821\u9a8c\u3002<\/li>\n<li>\u68c0\u7d22\u673a\u5236\u4e3a\u811a\u672c\u751f\u6210\u63d0\u4f9b\u624b\u518c\u548c\u89c4\u5219 grounding\u3002<\/li>\n<li>\u4e0d\u540c EDA \u5de5\u5177\u7248\u672c\u548c\u9879\u76ee flow \u5dee\u5f02\u5927\u3002<\/li>\n<li>compiler\/schema \u9700\u8981\u6301\u7eed\u7ef4\u62a4\u3002<\/li>\n<li>\u811a\u672c\u80fd\u8fd0\u884c\u4e0d\u4ee3\u8868\u8bbe\u8ba1 QoR \u6700\u4f18\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; JARVIS \u662f ChipNeMo \u811a\u672c\u751f\u6210\u7528\u4f8b\u7684\u5de5\u7a0b\u5316\u5347\u7ea7\u3002\u5b83\u8bf4\u660e EDA Copilot \u7684\u6b63\u786e\u59ff\u52bf\u4e0d\u662f\u8ba9\u6a21\u578b\u80cc\u6240\u6709 API&#xff0c;\u800c\u662f\u628a API\/\u89c4\u5219\u505a\u6210\u53ef\u68c0\u7d22\u3001\u53ef\u7f16\u8bd1\u3001\u53ef\u53cd\u9988\u7684\u7ea6\u675f\u7cfb\u7edf\u3002\u811a\u672c\u751f\u6210\u6bd4 RTL \u751f\u6210\u66f4\u5bb9\u6613\u77ed\u671f\u843d\u5730&#xff0c;\u56e0\u4e3a\u53cd\u9988\u66f4\u660e\u786e\u3001\u5931\u8d25\u6210\u672c\u66f4\u4f4e\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u5148\u505a\u5185\u90e8 Tcl\/Python EDA script assistant\u3002 \u5173\u952e\u8d44\u4ea7\u662f\u5de5\u5177\u547d\u4ee4 schema\u3001API compiler\u3001\u9519\u8bef\u4fee\u590d\u6837\u672c&#xff0c;\u800c\u4e0d\u662f UI \u804a\u5929\u7a97\u53e3\u3002 \u5c40\u9650\u662f\u4e0d\u540c EDA \u5de5\u5177\u3001\u7248\u672c\u3001license \u548c\u9879\u76ee flow \u5dee\u5f02\u4f1a\u5bfc\u81f4\u6cdb\u5316\u56f0\u96be\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>ChipNeMo \u00b7 Marco \u00b7 Timing Agent<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b custom compiler \u548c multi-agent feedback loop\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>ScaleRTL: Scaling LLMs with Reasoning Data and Test-Time Compute for Accurate RTL Code Generation<\/h4>\n<p>2025 \u00b7 MLCAD 2025 \/ arXiv \u00b7 AI4Design \u00b7 Reasoning \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04lvmqljuxz0t.png\" alt=\"Paper 14 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.1&#xff5c;Figure 1: Comparison of LLMs on RTL Coding benchmarks \u2014 ScaleRTL\u2020 refers to the test-time scaling variant of ScaleRTL.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-040fnccjvn20o.png\" alt=\"Paper 14 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.5&#xff5c;Table 2: Functional Correctness on RTL benchmarks \u2014 We highlight the top first, second, and third results per benchmark.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>RTL \u751f\u6210\u4e5f\u5b58\u5728 reasoning data \u548c test-time compute \u7684\u6269\u5c55\u6536\u76ca&#xff0c;\u4e0d\u80fd\u53ea\u9760\u666e\u901a\u4ee3\u7801\u5fae\u8c03\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u8f6f\u4ef6\u4ee3\u7801 benchmark \u4e0a reasoning model \u5df2\u7ecf\u663e\u793a\u4f18\u52bf&#xff0c;\u4f46 RTL \u751f\u6210\u53d7\u9650\u4e8e\u9ad8\u8d28\u91cf\u63a8\u7406\u6570\u636e\u7a00\u7f3a\u3002\u5df2\u6709 RTL \u5fae\u8c03\u6a21\u578b\u591a\u662f\u975e reasoning&#xff0c;\u6d4b\u8bd5\u65f6\u6269\u5c55\u80fd\u529b\u5f31&#xff0c;\u5931\u8d25\u540e\u7f3a\u5c11\u7ed3\u6784\u5316\u7ea0\u504f\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6574\u7406\u5927\u89c4\u6a21\u957f chain-of-thought RTL reasoning traces&#xff0c;\u5e73\u5747\u957f\u5ea6\u7ea6 56K tokens&#xff0c;\u603b\u91cf\u7ea6 3.5B tokens\u3002<\/li>\n<li>\u5bf9 reasoning model \u505a RTL \u65b9\u5411 SFT&#xff0c;\u4f7f\u5176\u5185\u5316\u786c\u4ef6\u8bed\u4e49\u548c\u8c03\u8bd5\u7b56\u7565\u3002<\/li>\n<li>\u5728\u63a8\u7406\u9636\u6bb5\u5f15\u5165 correction prompting \u548c test-time scaling\u3002<\/li>\n<li>\u5728 VerilogEval\u3001RTLLM \u7b49 benchmark \u4e0a\u4e0e 18 \u4e2a baseline \u6bd4\u8f83\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>RTL coding \u4e5f\u80fd\u4ece reasoning data \u548c test-time compute \u83b7\u76ca\u3002<\/li>\n<li>\u63a8\u7406\u8f68\u8ff9\u5fc5\u987b\u8fde\u63a5\u786c\u4ef6\u7ea6\u675f\u548c\u9a8c\u8bc1\u53cd\u9988&#xff0c;\u957f CoT \u672c\u8eab\u4e0d\u662f\u76ee\u6807\u3002<\/li>\n<li>\u8bad\u7ec3\u65f6\u6570\u636e\u6269\u5c55\u548c\u63a8\u7406\u65f6\u641c\u7d22\u662f\u4e24\u6761\u4e92\u8865\u6269\u5c55\u5f8b\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u6784\u9020\u5e73\u5747\u7ea6 56K tokens \u7684\u957f RTL reasoning traces\u3002<\/li>\n<li>\u603b\u6570\u636e\u89c4\u6a21\u7ea6 3.5B tokens\u3002<\/li>\n<li>\u5f15\u5165 correction prompting \u548c test-time scaling\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a\u5728 VerilogEval \u548c RTLLM \u4e0a\u8d85\u8fc7 18 \u4e2a baseline\u3002<\/li>\n<li>\u76f8\u5bf9\u63d0\u5347\u6700\u9ad8\u7ea6 18.4% \u548c 12.7%\u3002<\/li>\n<li>\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6709\u6536\u76ca\u4f46\u5448\u73b0\u9010\u6e10\u9971\u548c\u3002<\/li>\n<li>\u957f\u63a8\u7406\u6210\u672c\u9ad8&#xff0c;\u5fc5\u987b\u8bc4\u4f30 token cost \u548c latency\u3002<\/li>\n<li>\u63a8\u7406\u6570\u636e\u8d28\u91cf\u6bd4\u957f\u5ea6\u66f4\u91cd\u8981\u3002<\/li>\n<li>\u4ecd\u672a\u5b8c\u5168\u89e3\u51b3 repository-level \u591a\u6587\u4ef6\u4f9d\u8d56\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; ScaleRTL \u628a\u201c\u63a8\u7406\u6a21\u578b\u201d\u8def\u7ebf\u5e26\u5165 RTL\u3002\u5b83\u4e0e CraftRTL \u4e92\u8865&#xff1a;CraftRTL \u89e3\u51b3\u6570\u636e\u6b63\u786e\u6027\u548c\u975e\u6587\u672c\u89c4\u683c\u8986\u76d6&#xff0c;ScaleRTL \u89e3\u51b3\u957f\u63a8\u7406\u8f68\u8ff9\u548c\u6d4b\u8bd5\u65f6\u641c\u7d22\u3002\u771f\u6b63\u7684\u7ed3\u8bba\u4e0d\u662f CoT \u8d8a\u957f\u8d8a\u597d&#xff0c;\u800c\u662f reasoning trace \u8981\u548c\u53ef\u9a8c\u8bc1\u89c4\u5219\u3001\u5931\u8d25\u53cd\u9988\u7ed3\u5408\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u590d\u6742 RTL \u9898\u76ee\u3001\u534f\u8bae\u903b\u8f91\u3001FSM \u8bbe\u8ba1&#xff0c;\u4f46\u9700\u8981\u63a7\u5236\u63a8\u7406\u6210\u672c\u3002 test-time compute \u6709\u6536\u76ca\u4f46\u4f1a\u9971\u548c&#xff0c;\u5fc5\u987b\u6d4b pass&#064;k\u3001latency\u3001token cost\u3002 \u5c40\u9650\u662f\u957f CoT \u6570\u636e\u6784\u9020\u6602\u8d35&#xff0c;\u4e14\u672a\u5fc5\u76f4\u63a5\u8fc1\u79fb\u5230\u771f\u5b9e repo \u591a\u6587\u4ef6\u4efb\u52a1\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>CraftRTL \u00b7 ACE-RTL \u00b7 CVDP<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b reasoning trace \u6784\u9020\u548c test-time scaling \u66f2\u7ebf\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>PRO-V-R1: Reasoning Enhanced Programming Agent for RTL Verification<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4EDA \u00b7 Verification Agent \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04mhbmrdsmlrx.png\" alt=\"Paper 15 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.4&#xff5c;Figure 4: Overview of Agentic training framework.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04bsqfclnejjh.png\" alt=\"Paper 15 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.6&#xff5c;Table 3: Ablation Study of the PRO-V Framework on VerilogEval-v2. We progressively add our agentflow (PRO-V Flow), (SFT), and RL to the base model.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>RTL verification \u5e94\u88ab\u5efa\u6a21\u4e3a program-in-the-loop agent workflow&#xff0c;\u5c0f\u6a21\u578b\u7ecf\u8fc7 SFT\/GRPO \u4e5f\u80fd\u8d85\u8fc7\u95ed\u6e90\u5927\u6a21\u578b\u57fa\u7ebf\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>RTL verification \u6d88\u8017\u5927\u91cf\u5f00\u53d1\u65f6\u95f4\u3002\u73b0\u6709\u81ea\u52a8\u9a8c\u8bc1\u65b9\u6cd5\u5e38\u4f9d\u8d56 GPT-4o \u7b49\u95ed\u6e90\u5927\u6a21\u578b\u751f\u6210 Python reference \u6216 testbench&#xff0c;\u6210\u672c\u9ad8\u3001\u6570\u636e\u9690\u79c1\u98ce\u9669\u9ad8&#xff0c;\u5e76\u4e14\u7f3a\u5c11\u7aef\u5230\u7aef\u53ef\u8bad\u7ec3\u5f00\u6e90\u65b9\u6848\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6784\u5efa PRO-V system&#xff0c;\u5c06\u9a8c\u8bc1\u62c6\u6210 Python functional reference model\u3001testcase generation\u3001simulator execution \u548c feedback\u3002<\/li>\n<li>\u7528 simulation-validated expert trajectories \u505a SFT\u3002<\/li>\n<li>\u8bbe\u8ba1 verification-specific rewards&#xff0c;\u5e76\u7528 GRPO \u5f3a\u5316\u5b66\u4e60\u4f18\u5316 agent workflow\u3002<\/li>\n<li>\u8bc4\u4f30 functional correctness \u548c mutant robustness\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>RTL verification \u53ef\u4ee5\u88ab\u5efa\u6a21\u4e3a program-in-the-loop \u7684\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u3002<\/li>\n<li>\u5f00\u6e90\u5c0f\u6a21\u578b\u5982\u679c\u914d\u5408\u5de5\u5177\u3001\u8f68\u8ff9\u548c reward&#xff0c;\u80fd\u6311\u6218\u95ed\u6e90\u5927\u6a21\u578b\u3002<\/li>\n<li>\u9a8c\u8bc1 agent \u7684\u76ee\u6807\u4e0d\u662f\u5199\u6f02\u4eae testbench&#xff0c;\u800c\u662f\u6700\u5927\u5316\u529f\u80fd\u6b63\u786e\u548c fault detection\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u63d0\u51fa PRO-V system&#xff0c;\u628a FRM\u3001testcase\u3001simulator\u3001feedback \u4e32\u6210 agent workflow\u3002<\/li>\n<li>\u7528 simulation-validated expert trajectories \u505a SFT\u3002<\/li>\n<li>\u7528 GRPO \u548c verification-specific rewards \u505a RL\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a 8B PRO-V-R1 \u8fbe\u5230 57.7% functional correctness\u3002<\/li>\n<li>full-mutant robustness \u8fbe\u5230 34.0%&#xff0c;\u8d85\u8fc7 GPT-4o \u5bf9\u7167\u3002<\/li>\n<li>\u76f8\u6bd4 CorrectBench \u901f\u5ea6\u4e5f\u6709\u660e\u663e\u4f18\u52bf\u3002<\/li>\n<li>Python FRM \u4e0e RTL \u8bed\u4e49\u4e0d\u4e00\u81f4\u4f1a\u5f15\u5165\u4f2a oracle\u3002<\/li>\n<li>mutant robustness \u4e0d\u7b49\u4e8e\u5b8c\u6574\u8986\u76d6\u7387\u6536\u655b\u3002<\/li>\n<li>\u4ecd\u9700\u4eba\u5de5\u5236\u5b9a\u9a8c\u8bc1\u8ba1\u5212\u548c signoff \u6807\u51c6\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; PRO-V-R1 \u7684\u610f\u4e49\u5728\u4e8e\u628a\u9a8c\u8bc1\u4ee3\u7406\u4ece prompt \u5de5\u7a0b\u63a8\u8fdb\u5230\u53ef\u8bad\u7ec3\u7cfb\u7edf\u3002\u5b83\u4e0d\u53ea\u662f\u751f\u6210 testbench&#xff0c;\u800c\u662f\u628a\u6a21\u578b\u3001\u7a0b\u5e8f\u5de5\u5177\u3001\u4eff\u771f\u5668\u548c reward \u7ec4\u7ec7\u6210\u95ed\u73af\u3002\u5bf9\u4f01\u4e1a\u800c\u8a00&#xff0c;\u8fd9\u4ee3\u8868\u4e00\u79cd\u53ef\u79c1\u6709\u5316\u3001\u5c0f\u6a21\u578b\u5316\u7684\u9a8c\u8bc1\u81ea\u52a8\u5316\u65b9\u5411\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u505a testbench assistant\u3001reference model generator\u3001mutant-based robustness checker\u3002 \u9700\u8981\u4fdd\u6301 Verilog\/Python \u8bed\u4e49\u4e00\u81f4&#xff0c;\u907f\u514d reference model \u81ea\u8eab\u5f15\u5165\u9519\u8bef\u3002 \u5c40\u9650\u662f\u9a8c\u8bc1\u8986\u76d6\u8d28\u91cf\u4e0d\u7b49\u4e8e\u53d1\u73b0\u6240\u6709 bug&#xff0c;\u4ecd\u4e0d\u80fd\u66ff\u4ee3\u4eba\u5de5 signoff \u548c\u8986\u76d6\u7387\u7b56\u7565\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>FVEval \u00b7 AssertionForge \u00b7 CVDP<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b PRO-V system\u3001trajectory construction \u548c reward design\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Comprehensive Verilog Design Problems: A Next-Generation Benchmark Dataset<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4Design \u00b7 Benchmark \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04gfxlfhzau1r.png\" alt=\"Paper 16 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.4&#xff5c;Figure 1: Benchmark Evaluation Flow.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04kdaksgkshzj.png\" alt=\"Paper 16 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.15&#xff5c;Detected table region by pdfplumber fallback, page 15, table 1, rows 56.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u77ed\u5c0f RTL benchmark \u5df2\u4e0d\u8db3\u4ee5\u533a\u5206 agent \u80fd\u529b&#xff0c;CVDP \u7528 783 \u4e2a\u4e13\u5bb6\u95ee\u9898\u8986\u76d6\u66f4\u771f\u5b9e\u7684\u8bbe\u8ba1\u3001\u9a8c\u8bc1\u3001\u8c03\u8bd5\u548c\u6280\u672f\u95ee\u7b54\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>VerilogEval \u7c7b\u77ed\u9898\u9010\u6e10\u9971\u548c&#xff0c;\u65e0\u6cd5\u5145\u5206\u8bc4\u4f30\u771f\u5b9e\u786c\u4ef6\u5de5\u7a0b\u4e2d\u7684\u591a\u6587\u4ef6\u4f9d\u8d56\u3001\u6a21\u5757\u590d\u7528\u3001\u9a8c\u8bc1\u3001lint\/QoR\u3001\u6280\u672f\u95ee\u7b54\u548c\u5de5\u5177\u4ea4\u4e92\u3002agent \u5982\u679c\u53ea\u4f1a\u5355\u6587\u4ef6\u751f\u6210&#xff0c;\u79bb\u771f\u5b9e\u5de5\u7a0b\u8fd8\u5f88\u8fdc\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6784\u5efa 783 \u4e2a\u7531\u786c\u4ef6\u4e13\u5bb6\u7f16\u5199\u7684\u95ee\u9898&#xff0c;\u8986\u76d6 13 \u4e2a\u4efb\u52a1\u7c7b\u522b\u3002<\/li>\n<li>\u4efb\u52a1\u5206\u4e3a non-agentic code generation\u3001code comprehension \u548c agentic code generation\u3002<\/li>\n<li>\u63d0\u4f9b Dockerized agent environment\u3001test harness \u548c\u5de5\u5177\u4ea4\u4e92\u57fa\u7840\u8bbe\u65bd\u3002<\/li>\n<li>\u7528\u53ef\u6267\u884c\u6d4b\u8bd5\u3001BLEU\u3001LLM judge \u7b49\u7ec4\u5408\u65b9\u5f0f\u8bc4\u4f30\u4e0d\u540c\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u4e0b\u4e00\u4ee3 benchmark \u5fc5\u987b\u4ece\u77ed\u9898\u8d70\u5411 repository\/context\/tool-heavy \u4efb\u52a1\u3002<\/li>\n<li>agent \u80fd\u529b\u5e94\u5728\u8bbe\u8ba1\u3001\u9a8c\u8bc1\u3001\u7406\u89e3\u3001\u590d\u7528\u548c\u4fee\u590d\u591a\u4e2a\u7ef4\u5ea6\u540c\u65f6\u8bc4\u4f30\u3002<\/li>\n<li>\u9690\u85cf\u6d4b\u8bd5\u548c Dockerized environment \u662f\u9632\u6b62\u8fc7\u62df\u5408\u4e0e\u8bc4\u4f30\u6cc4\u9732\u7684\u5173\u952e\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>783 \u4e2a\u786c\u4ef6\u4e13\u5bb6\u7f16\u5199\u7684\u95ee\u9898\u3002<\/li>\n<li>\u8986\u76d6 13 \u4e2a\u4efb\u52a1\u7c7b\u522b&#xff0c;\u542b non-agentic \u4e0e agentic \u683c\u5f0f\u3002<\/li>\n<li>\u63d0\u4f9b Dockerized agent environment \u548c test harness\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a SOTA \u6a21\u578b\u5728 code generation \u4e0a pass&#064;1 \u4e0d\u8d85\u8fc7\u7ea6 34%\u3002<\/li>\n<li>agentic tasks \u5c24\u5176 RTL reuse \u4e0e verification \u66f4\u56f0\u96be\u3002<\/li>\n<li>CVDP \u540e\u7eed\u76f4\u63a5\u9a71\u52a8 ACE-RTL\u3001Trace2Skill \u7b49\u5de5\u4f5c\u3002<\/li>\n<li>benchmark \u4ecd\u65e0\u6cd5\u5b8c\u5168\u4ee3\u8868\u4f01\u4e1a\u79c1\u6709\u5de5\u5177\u94fe\u548c IP \u4f9d\u8d56\u3002<\/li>\n<li>LLM judge \u7528\u4e8e\u7406\u89e3\u4efb\u52a1\u65f6\u9700\u8981\u8c28\u614e\u3002<\/li>\n<li>\u4efb\u52a1\u6269\u5c55\u548c hidden verifier \u7ef4\u62a4\u6210\u672c\u9ad8\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; CVDP \u662f NVIDIA \u8def\u7ebf\u4ece benchmark \u5230\u771f\u5b9e\u5de5\u7a0b\u7684\u5206\u6c34\u5cad\u3002\u5b83\u628a\u95ee\u9898\u4ece\u201c\u5199\u4e00\u4e2a\u6a21\u5757\u201d\u63a8\u8fdb\u5230\u201c\u5728\u5de5\u7a0b\u4e0a\u4e0b\u6587\u91cc\u5b9a\u4f4d\u3001\u4fee\u6539\u3001\u9a8c\u8bc1\u201d\u3002\u540e\u7eed ACE-RTL \u548c Trace2Skill \u57fa\u672c\u90fd\u56f4\u7ed5 CVDP \u66b4\u9732\u51fa\u7684\u957f\u4e0a\u4e0b\u6587\u56f0\u96be\u7ee7\u7eed\u63a8\u8fdb\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u4f01\u4e1a\u5185\u90e8 benchmark \u5e94\u5b66\u4e60 CVDP&#xff1a;\u9898\u76ee\u8981\u6765\u81ea\u771f\u5b9e\u5931\u8d25\u6a21\u5f0f&#xff0c;\u800c\u4e0d\u662f\u73a9\u5177\u9898\u3002 \u8981\u540c\u65f6\u6d4b\u8bbe\u8ba1\u3001\u9a8c\u8bc1\u3001\u7406\u89e3\u3001\u811a\u672c\u3001\u590d\u7528\u3001\u4fee\u590d&#xff0c;\u800c\u4e0d\u662f\u53ea\u6d4b RTL \u751f\u6210\u3002 \u5c40\u9650\u662f CVDP \u4ecd\u662f benchmark&#xff0c;\u771f\u5b9e IP \u7684\u4fdd\u5bc6\u4f9d\u8d56\u3001license\u3001\u5de5\u5177\u7248\u672c\u4f1a\u66f4\u590d\u6742\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>VerilogEval \u00b7 ACE-RTL \u00b7 Trace2Skill<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b 13 \u7c7b\u4efb\u52a1\u3001agentic \u73af\u5883\u548c pass&#064;1 headroom\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Multimodal Chip Physical Design Engineer Assistant<\/h4>\n<p>2025 \u00b7 arXiv \u00b7 AI4EDA \u00b7 Multimodal \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04yljqilvnv2w.png\" alt=\"Paper 17 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.9&#xff5c;Figure 5: By following the top model-attributed features and adjusting actionable design parameters, the resulting design exhibits significantly lower predicted and actual congestion.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-043b33f1yrzxs.png\" alt=\"Paper 17 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.7&#xff5c;Table 1: Congestion prediction performance on CircuitNet across pixel-based and correlation-based metrics.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u7269\u7406\u8bbe\u8ba1\u52a9\u624b\u4e0d\u80fd\u53ea\u8f93\u51fa\u62e5\u585e\u70ed\u529b\u56fe&#xff0c;\u8fd8\u8981\u628a\u56fe\u50cf\u3001\u8868\u683c\u3001\u56fe\u7ed3\u6784\u8f6c\u6210\u5de5\u7a0b\u5e08\u53ef\u91c7\u7eb3\u7684\u8bbe\u8ba1\u5efa\u8bae\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u540e\u7aef\u7269\u7406\u8bbe\u8ba1\u4f9d\u8d56 layout image\u3001routing congestion\u3001pin distribution\u3001cluster\u3001graph\u3001tabular features \u7b49\u591a\u6a21\u6001\u4fe1\u606f\u3002\u4f20\u7edf\u6a21\u578b\u53ef\u80fd\u80fd\u9884\u6d4b\u62e5\u585e&#xff0c;\u4f46\u5f88\u96be\u89e3\u91ca\u4e3a\u4ec0\u4e48\u62e5\u585e\u3001\u600e\u4e48\u6539\u3001\u6539\u52a8\u5bf9 tradeoff \u7684\u5f71\u54cd\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6784\u5efa Multimodal LLM Assistant&#xff0c;\u878d\u5408 visual\u3001tabular\u3001circuit graph \u7b49\u8f93\u5165\u3002<\/li>\n<li>\u7528 MLLM-guided genetic prompting \u81ea\u52a8\u751f\u6210\u548c\u7b5b\u9009\u7269\u7406\u8bbe\u8ba1\u7279\u5f81\u3002<\/li>\n<li>\u7528 interpretable preference learning \u5efa\u6a21\u62e5\u585e\u76f8\u5173 tradeoff\u3002<\/li>\n<li>\u8f93\u51fa Design Suggestion Deck&#xff0c;\u628a\u9884\u6d4b\u8f6c\u5316\u4e3a\u53ef\u89e3\u91ca\u7684\u4f18\u5316\u52a8\u4f5c\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u7269\u7406\u8bbe\u8ba1 agent \u5fc5\u987b\u5904\u7406\u89c6\u89c9\u3001\u8868\u683c\u548c\u56fe\u7ed3\u6784&#xff0c;\u4e0d\u662f\u7eaf\u6587\u672c\u6a21\u578b\u80fd\u5b8c\u5168\u8986\u76d6\u3002<\/li>\n<li>\u53ef\u89e3\u91ca\u5efa\u8bae\u6bd4\u5355\u4e00\u9884\u6d4b\u70ed\u56fe\u66f4\u63a5\u8fd1\u5de5\u7a0b\u4ef7\u503c\u3002<\/li>\n<li>\u591a\u6a21\u6001 feature generation \u53ef\u4ee5\u628a\u7248\u56fe\u611f\u77e5\u8f6c\u6362\u4e3a\u53ef\u64cd\u4f5c\u8bbe\u8ba1\u52a8\u4f5c\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u878d\u5408 routability image\u3001tabular features\u3001circuit graph\u3002<\/li>\n<li>\u7528 MLLM-guided genetic prompting \u81ea\u52a8\u751f\u6210\u7279\u5f81\u3002<\/li>\n<li>\u8f93\u51fa Design Suggestion Deck&#xff0c;\u800c\u4e0d\u53ea\u662f congestion score\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u5728 CircuitNet benchmark \u4e0a\u62a5\u544a\u7cbe\u5ea6\u548c explainability \u5747\u4f18\u4e8e\u65e2\u6709\u65b9\u6cd5\u3002<\/li>\n<li>case studies \u8868\u660e\u5efa\u8bae\u4e0e\u5de5\u7a0b\u76f4\u89c9\u4e00\u81f4\u3002<\/li>\n<li>\u7ed3\u5408 preference learning \u5efa\u6a21\u62e5\u585e\u76f8\u5173 tradeoff\u3002<\/li>\n<li>\u53ef\u89e3\u91ca\u5efa\u8bae\u4ecd\u9700\u7269\u7406\u8bbe\u8ba1\u4e13\u5bb6\u5ba1\u67e5\u3002<\/li>\n<li>\u516c\u5f00 CircuitNet \u4e0e\u771f\u5b9e\u5148\u8fdb\u8282\u70b9\u8bbe\u8ba1\u5206\u5e03\u4e0d\u540c\u3002<\/li>\n<li>\u4ece\u5efa\u8bae\u5230\u81ea\u52a8 ECO \u8fd8\u9700\u8981\u95ed\u73af\u9a8c\u8bc1\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; \u8fd9\u7bc7\u8bba\u6587\u62d3\u5bbd\u4e86 NVIDIA AI4EDA \u7684\u8f93\u5165\u5f62\u6001\u3002\u540e\u7aef\u8bbe\u8ba1\u4e2d\u7684\u5173\u952e\u8bc1\u636e\u5f88\u591a\u4e0d\u662f\u6587\u672c&#xff0c;\u800c\u662f\u70ed\u56fe\u3001\u7248\u56fe\u3001\u8868\u683c\u3001\u8def\u5f84\u548c\u56fe\u7ed3\u6784\u3002\u672a\u6765\u7269\u7406\u8bbe\u8ba1 agent \u5fc5\u987b\u80fd\u770b\u56fe\u3001\u8bfb\u8868\u3001\u8bfb report&#xff0c;\u5e76\u628a\u8fd9\u4e9b\u4fe1\u53f7\u5f52\u7eb3\u6210\u5de5\u7a0b\u5efa\u8bae\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u62e5\u585e\u89e3\u91ca\u3001floorplan review\u3001pin placement review\u3001QoR summary \u7b49\u573a\u666f\u3002 \u5bf9\u5185\u90e8\u62a5\u544a\u7cfb\u7edf\u5f88\u6709\u4ef7\u503c&#xff1a;\u628a\u56fe\u50cf\u6307\u6807\u548c\u6587\u5b57\u89e3\u91ca\u7ed1\u5b9a&#xff0c;\u800c\u4e0d\u662f\u53ea\u7ed9\u70ed\u56fe\u3002 \u5c40\u9650\u662f\u591a\u6a21\u6001\u89e3\u91ca\u7684\u53ef\u4fe1\u5ea6\u9700\u8981\u4e13\u5bb6\u5ba1\u67e5&#xff0c;\u5efa\u8bae\u4e0d\u7b49\u4e8e\u81ea\u52a8 ECO\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>CircuitNet \u00b7 ForgeEDA \u00b7 AiEDA<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b Genetic Instruct\u3001preference learning \u548c Design Suggestion Deck\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>ACE-RTL: When Agentic Context Evolution Meets RTL-Specialized LLMs<\/h4>\n<p>2026 \u00b7 arXiv \u00b7 AI4Design \u00b7 Context Evolution \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04potffto4h3g.png\" alt=\"Paper 18 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.3&#xff5c;Figure 2: Overview of ACE-RTL and its parallel scaling strategy.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04e23y2vm03ml.png\" alt=\"Paper 18 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.4&#xff5c;Table 1 shows prior RTL LLMs struggle on CVDP despite performing well on earlier RTL benchmarks, suggesting limited<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>RTL-specialized model \u548c frontier agent \u5404\u6709\u4f18\u52bf&#xff0c;Agentic Context Evolution \u80fd\u628a\u9886\u57df\u6a21\u578b\u653e\u8fdb\u8fed\u4ee3\u4fee\u590d\u95ed\u73af\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>\u4e24\u6761\u8def\u7ebf\u957f\u671f\u5206\u79bb&#xff1a;\u9886\u57df\u6a21\u578b\u61c2 RTL \u4f46 agentic \u534f\u4f5c\u5f31&#xff1b;\u901a\u7528 frontier agent \u5de5\u5177\u4f7f\u7528\u5f3a\u4f46\u786c\u4ef6\u8bed\u4e49\u4e0d\u4e00\u5b9a\u7a33\u3002\u771f\u5b9e CVDP \u4efb\u52a1\u9700\u8981\u4e24\u8005\u7ed3\u5408&#xff1a;\u65e2\u61c2 RTL&#xff0c;\u53c8\u80fd\u8bfb\u53cd\u9988\u3001\u6539\u4e0a\u4e0b\u6587\u3001\u91cd\u8bd5\u548c\u534f\u8c03\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u8bad\u7ec3 RTL-specialized generator&#xff0c;\u6570\u636e\u89c4\u6a21\u5305\u62ec\u7ea6 1.7M RTL samples\u3002<\/li>\n<li>\u63d0\u51fa Generator\u3001Reflector\u3001Coordinator \u4e09\u7ec4\u4ef6\u3002<\/li>\n<li>\u901a\u8fc7 Agentic Context Evolution \u6301\u7eed\u6539\u5199\u4e0a\u4e0b\u6587\u548c\u4fee\u590d\u65b9\u5411\u3002<\/li>\n<li>\u4f7f\u7528 parallel scaling \u5e76\u884c\u63a2\u7d22\u591a\u4e2a\u8c03\u8bd5\u8f68\u8ff9&#xff0c;\u7f29\u77ed first success time\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u9886\u57df\u6a21\u578b\u548c agent loop \u4e0d\u5e94\u5206\u79bb&#xff1a;\u9886\u57df\u6a21\u578b\u53ef\u4ee5\u4f5c\u4e3a\u8fed\u4ee3\u4fee\u590d\u6838\u5fc3\u3002<\/li>\n<li>\u4e0a\u4e0b\u6587\u662f\u53ef\u4f18\u5316\u72b6\u6001&#xff0c;\u4e0d\u53ea\u662f prompt \u6587\u672c\u3002<\/li>\n<li>\u5e76\u884c\u8c03\u8bd5\u8f68\u8ff9\u80fd\u7528\u8ba1\u7b97\u6362 first success&#xff0c;\u4f46\u5fc5\u987b\u6709 verifier \u7ba1\u63a7\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u628a RTL-specialized LLM \u5d4c\u5165 Agentic Context Evolution\u3002<\/li>\n<li>\u4f7f\u7528 Generator\u3001Reflector\u3001Coordinator \u4e09\u7ec4\u4ef6\u3002<\/li>\n<li>\u5728 CVDP \u4e0a\u5c55\u793a\u76f8\u5bf9 baseline \u7684\u663e\u8457 pass-rate improvement\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u8bba\u6587\u62a5\u544a ACE-RTL \u5728 CVDP \u4e0a\u6700\u9ad8\u5e26\u6765\u7ea6 41.02% pass-rate improvement\u3002<\/li>\n<li>RTL generator \u8bad\u7ec3\u6570\u636e\u7ea6 1.7M samples\u3002<\/li>\n<li>\u7ed3\u8bba\u6307\u51fa future direction \u5305\u62ec\u81ea\u52a8 testbench \u751f\u6210\u548c waveform-enhanced reasoning\u3002<\/li>\n<li>\u5e76\u884c scaling \u6210\u672c\u9ad8\u3002<\/li>\n<li>\u9700\u8981\u9ad8\u8d28\u91cf\u4eff\u771f\/\u9690\u85cf\u6d4b\u8bd5\u53cd\u9988&#xff0c;\u5426\u5219\u4e0a\u4e0b\u6587\u6f14\u5316\u4f1a\u6f02\u79fb\u3002<\/li>\n<li>\u9886\u57df\u6a21\u578b\u53ef\u80fd\u8fc7\u62df\u5408 benchmark \u98ce\u683c\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; ACE-RTL \u7684\u5173\u952e\u89c2\u70b9\u662f&#xff1a;\u9886\u57df\u6a21\u578b\u4e0d\u5e94\u53ea\u4f5c\u4e3a\u4e00\u6b21\u6027\u751f\u6210\u5668&#xff0c;\u800c\u5e94\u4f5c\u4e3a agent loop \u7684\u6838\u5fc3\u6267\u884c\u5668\u3002\u4e0a\u4e0b\u6587\u672c\u8eab\u6210\u4e3a\u53ef\u4f18\u5316\u5bf9\u8c61&#xff0c;\u6a21\u578b\u770b\u5230\u4ec0\u4e48\u3001\u5982\u4f55\u7ec4\u7ec7\u5931\u8d25\u53cd\u9988\u3001\u4f55\u65f6\u91cd\u542f\u641c\u7d22\u65b9\u5411&#xff0c;\u90fd\u4f1a\u5f71\u54cd\u6700\u7ec8 pass rate\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u5185\u90e8\u5c06\u4e13\u7528 RTL \u6a21\u578b\u5d4c\u5165 agent&#xff0c;\u800c\u4e0d\u662f\u53ea\u62ff\u5b83\u505a\u8865\u5168\u3002 \u8981\u8bb0\u5f55\u6bcf\u8f6e\u4e0a\u4e0b\u6587\u3001\u53cd\u9988\u3001\u4fee\u6539\u548c\u7ed3\u679c&#xff0c;\u624d\u80fd\u5206\u6790\u6f14\u5316\u662f\u5426\u6709\u6548\u3002 \u5c40\u9650\u662f\u5e76\u884c\u641c\u7d22\u6210\u672c\u9ad8&#xff0c;\u4e14\u4ecd\u4f9d\u8d56\u9ad8\u8d28\u91cf testbench\/verifier\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>ScaleRTL \u00b7 CVDP \u00b7 Trace2Skill<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b Generator\/Reflector\/Coordinator \u5982\u4f55\u66f4\u65b0\u4e0a\u4e0b\u6587\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Trace2Skill: Verifier-Guided Skill Evolution for Long-Context EDA Agents<\/h4>\n<p>2026 \u00b7 arXiv \u00b7 AI4Design \u00b7 Skill Evolution \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04mtuarirdnr2.png\" alt=\"Paper 19 \u8bba\u6587\u5185\u56fe\" \/><\/p>\n<p>\u8bba\u6587\u5185\u56fe&#xff5c;PDF p.4&#xff5c;Figure 2: AgentQ proxy on 96 completed OSS seed-skill base- line rollouts, grouped by hidden verifier outcome and colored by CID. Each point is one rollout.<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04rmjvt35kuls.png\" alt=\"Paper 19 \u8bba\u6587\u5185\u8868\" \/><\/p>\n<p>\u8bba\u6587\u5185\u8868&#xff5c;PDF p.4&#xff5c;Table 2: Matched public OSS CVDP baseline results by CID. The subset contains public long-context agentic CVDP tasks whose harnesses do not require commercial EDA tools, with six sampled tasks per OSS-runnable CID.<\/p>\n<h5>\u8bba\u6587\u547d\u9898<\/h5>\n<p>\u4e0d\u6539\u6a21\u578b\u6743\u91cd&#xff0c;\u4e5f\u80fd\u628a\u6267\u884c\u8f68\u8ff9\u548c dense verifier feedback \u8f6c\u5316\u4e3a\u53ef\u6f14\u5316\u7684\u81ea\u7136\u8bed\u8a00 skill&#xff0c;\u4ece\u800c\u63d0\u5347\u957f\u4e0a\u4e0b\u6587 EDA agent\u3002<\/p>\n<h5>\u8981\u89e3\u51b3\u7684\u95ee\u9898<\/h5>\n<p>CVDP \u8fd9\u7c7b\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u5931\u8d25\u65f6&#xff0c;hidden verifier \u53ea\u7ed9 sparse pass\/fail&#xff0c;agent \u5f88\u96be\u5b66\u4e60\u3002\u5931\u8d25\u539f\u56e0\u53ef\u80fd\u662f include path\u3001testbench\u3001\u5c40\u90e8 RTL\u3001\u6784\u5efa\u4f9d\u8d56\u3001\u7f16\u8f91\u4f4d\u7f6e\u6216\u9a8c\u8bc1\u7b56\u7565\u9519\u8bef&#xff1b;\u5355\u7eaf\u591a\u91c7\u6837\u4e0d\u80fd\u7cfb\u7edf\u6027\u6539\u8fdb\u3002<\/p>\n<h5>\u65b9\u6cd5\u62c6\u89e3<\/h5>\n<ul>\n<li>\u56fa\u5b9a seed CVDP agent&#xff0c;\u4e0d\u505a RTL-specialized fine-tuning\u3002<\/li>\n<li>\u53cd\u590d\u8fd0\u884c\u4efb\u52a1\u5e76\u6316\u6398\u6210\u529f\/\u5931\u8d25 rollout traces\u3002<\/li>\n<li>\u5f15\u5165 dense verifier feedback&#xff0c;\u5728\u4e0d\u6cc4\u9732 hidden harness \u7684\u60c5\u51b5\u4e0b\u8fd4\u56de\u66f4\u7ec6\u7684\u8bca\u65ad\u4fe1\u53f7\u3002<\/li>\n<li>\u7528 oracle-mutator-selector \u5faa\u73af\u6f14\u5316\u81ea\u7136\u8bed\u8a00 skill&#xff0c;\u5e76\u7528 SkillQ\u3001AgentProgressQ\u3001SelectQ \u8bc4\u4f30\u3002<\/li>\n<\/ul>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<ul>\n<li>\u81ea\u7136\u8bed\u8a00 skill \u53ef\u4ee5\u4f5c\u4e3a\u53ef\u6f14\u5316 policy&#xff0c;\u6bd4\u6743\u91cd\u66f4\u65b0\u66f4\u53ef\u5ba1\u8ba1\u3002<\/li>\n<li>dense verifier feedback \u80fd\u628a\u7a00\u758f pass\/fail \u8f6c\u5316\u4e3a\u53ef\u5b66\u4e60\u4fe1\u53f7\u3002<\/li>\n<li>\u5931\u8d25 trace \u662f\u8d44\u4ea7&#xff0c;\u4e0d\u662f\u5e9f\u5f03\u65e5\u5fd7\u3002<\/li>\n<\/ul>\n<h5>\u5173\u952e\u4eae\u70b9<\/h5>\n<ul>\n<li>\u56fa\u5b9a seed CVDP agent&#xff0c;\u4e0d\u505a RTL fine-tuning\u3002<\/li>\n<li>\u7528 oracle-mutator-selector \u8fdb\u5316\u4efb\u52a1\u4e13\u7528 skill\u3002<\/li>\n<li>\u8bbe\u8ba1 SkillQ\u3001AgentProgressQ\u3001SelectQ \u8bc4\u4f30 skill \u4e0e agent \u5171\u540c\u8fdb\u6b65\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe\u4e0e\u8fb9\u754c<\/h5>\n<ul>\n<li>\u5728 8 \u4e2a seed agent \u5168\u5931\u8d25\u4efb\u52a1\u4e0a&#xff0c;\u5b8c\u6574 Trace2Skill &#043; dense feedback \u89e3\u51fa 6\/8\u3002<\/li>\n<li>hidden-verifier pass rate \u8fbe\u5230 33.6%&#xff0c;\u5bf9\u6bd4 seed agent 0\/8\u3002<\/li>\n<li>dense feedback alone \u4e0e sparse skill evolution \u5747\u4e0d\u5982\u5b8c\u6574\u914d\u7f6e\u3002<\/li>\n<li>dense verifier \u8bbe\u8ba1\u8fc7\u5f3a\u4f1a\u6cc4\u9732\u7b54\u6848&#xff0c;\u8fc7\u5f31\u53c8\u5b66\u4e0d\u5230\u3002<\/li>\n<li>skill \u53ef\u80fd\u8fc7\u62df\u5408\u5355\u4e00\u4efb\u52a1&#xff0c;\u9700\u8981\u8de8\u4efb\u52a1\u6cdb\u5316\u8bc4\u4f30\u3002<\/li>\n<li>\u9700\u8981 trace \u5b58\u50a8\u3001skill \u7248\u672c\u3001\u56de\u6eda\u548c\u5ba1\u8ba1\u57fa\u7840\u8bbe\u65bd\u3002<\/li>\n<\/ul>\n<p>\u89e3\u8bfb&#xff1a; Trace2Skill \u662f NVIDIA \u8def\u7ebf\u76ee\u524d\u6700\u63a5\u8fd1\u201c\u81ea\u8fdb\u5316\u5de5\u7a0b\u4ee3\u7406\u201d\u7684\u5de5\u4f5c\u3002\u5b83\u628a\u5931\u8d25\u8f68\u8ff9\u5f53\u6210\u8bad\u7ec3\u5916\u4f18\u5316\u6570\u636e&#xff0c;\u628a skill \u6587\u6863\u5f53\u6210 agent policy \u7684\u663e\u5f0f\u8f7d\u4f53\u3002\u4e0e\u9ed1\u76d2\u6743\u91cd\u66f4\u65b0\u76f8\u6bd4&#xff0c;skill \u6f14\u5316\u53ef\u8bfb\u3001\u53ef\u5ba1\u8ba1\u3001\u53ef\u56de\u6eda&#xff0c;\u66f4\u7b26\u5408 EDA \u5de5\u7a0b\u9700\u8981\u3002<\/p>\n<table>\n<tr>\u5de5\u7a0b\u4ef7\u503c\u9002\u5408\u5185\u90e8 agent \u5e73\u53f0\u6c89\u6dc0\u7ecf\u9a8c&#xff1a;\u6bcf\u6b21\u5931\u8d25\u90fd\u5e94\u8f6c\u5316\u4e3a\u53ef\u68c0\u7d22 skill \u6216\u89c4\u5219\u3002 \u9700\u8981\u5b8c\u5584 trace store\u3001verifier feedback schema\u3001skill \u7248\u672c\u7ba1\u7406\u548c\u9009\u62e9\u5668\u3002 \u5c40\u9650\u662f dense verifier \u7684\u8bbe\u8ba1\u5f88\u5173\u952e&#xff1b;\u53cd\u9988\u8fc7\u5c11\u5b66\u4e0d\u5230&#xff0c;\u53cd\u9988\u8fc7\u591a\u53ef\u80fd\u6cc4\u9732\u7b54\u6848\u6216\u8fc7\u62df\u5408\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5173\u8054\u8109\u7edc<\/td>\n<td>CVDP \u00b7 ACE-RTL \u00b7 Reflexion \u00b7 Voyager \u00b7 SWE-agent<\/td>\n<\/tr>\n<tr>\n<td>\u9605\u8bfb\u91cd\u70b9<\/td>\n<td>\u91cd\u70b9\u770b dense feedback\u3001skill evolution loop \u548c\u4e09\u4e2a\u8d28\u91cf\u6307\u6807\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u5168\u7403\u6a2a\u8f74&#xff1a;\u975e NVIDIA \u8bba\u6587\u3001\u5f00\u6e90\u9879\u76ee\u3001\u673a\u6784\u8def\u7ebf\u56fe\u4e0e\u4ea7\u4e1a\u5e73\u53f0\u9010\u9879\u6df1\u8bfb<\/h3>\n<p>\u5982\u679c\u8bf4 NVIDIA \u8fd9\u7ec4\u8bba\u6587\u63ed\u793a\u4e86 LLM\/agent \u8fdb\u5165\u82af\u7247\u5de5\u7a0b\u7684\u5fae\u89c2\u673a\u5236&#xff0c;\u90a3\u4e48\u5168\u7403\u6a2a\u8f74\u5c55\u793a\u7684\u662f\u540c\u4e00\u95ee\u9898\u5728\u5176\u4ed6\u8bba\u6587\u3001\u5f00\u6e90\u57fa\u7840\u8bbe\u65bd\u3001\u673a\u6784\u8def\u7ebf\u56fe\u548c\u5546\u4e1a EDA \u5e73\u53f0\u4e2d\u7684\u4e0d\u540c\u89e3\u6cd5\u3002\u672c\u8282\u4e0d\u662f\u9644\u5f55&#xff0c;\u800c\u662f\u4e0e NVIDIA \u7eb5\u8f74\u540c\u7b49\u91cd\u8981\u7684\u4e3b\u4f53&#xff1a;\u6bcf\u4e00\u9879\u90fd\u7ed9\u51fa\u6765\u6e90\u622a\u56fe\u3001\u95ee\u9898\u5b9a\u4e49\u3001\u65b9\u6cd5\u67b6\u6784\u3001\u8bc1\u636e\u94fe\u3001\u7406\u8bba\u7ed3\u8bba\u3001\u5de5\u7a0b\u542f\u793a\u548c\u9002\u7528\u8fb9\u754c\u3002<\/p>\n<p>\u5b66\u672f\u8bba\u6587 \/ \u5f3a\u5316\u5b66\u4e60\u5e03\u5c40<\/p>\n<h4>Google \/ DeepMind AlphaChip \u524d\u8eab&#xff1a;Chip Placement with Deep Reinforcement Learning<\/h4>\n<p>arXiv 2004.10746&#xff0c;DeepMind\/Google Research \u5b8f\u5355\u5143\u5e03\u5c40\u8def\u7ebf\u7684\u516c\u5f00\u8bba\u6587\u7248\u672c\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04p12ynkazdri.png\" alt=\"Google \/ DeepMind AlphaChip \u524d\u8eab\uff1aChip Placement with Deep Reinforcement Learning - \u7b56\u7565\/\u4ef7\u503c\u7f51\u7edc\u67b6\u6784\u622a\u56fe\" \/><\/p>\n<p>\u7b56\u7565\/\u4ef7\u503c\u7f51\u7edc\u67b6\u6784\u622a\u56fe \u00b7 PDF p.7 \u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a policy\/value network \u5982\u4f55\u63a5\u6536 netlist \u56fe\u3001\u5f53\u524d\u5b8f\u5355\u5143\u548c\u5143\u6570\u636e\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04cezqhewm0yb.png\" alt=\"Google \/ DeepMind AlphaChip \u524d\u8eab\uff1aChip Placement with Deep Reinforcement Learning - \u5b9e\u9a8c\u5bf9\u6bd4\u8868\u622a\u56fe\" \/><\/p>\n<p>\u5b9e\u9a8c\u5bf9\u6bd4\u8868\u622a\u56fe \u00b7 PDF p.11 \u8bba\u6587\u5185\u90e8\u8868\u683c\u5bf9\u6bd4 RL\u3001\u6a21\u62df\u9000\u706b\u548c\u4eba\u5de5\/\u4f20\u7edf\u65b9\u6848&#xff0c;\u662f\u5224\u65ad\u5176\u5de5\u7a0b\u610f\u4e49\u7684\u6838\u5fc3\u8bc1\u636e\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u5b8f\u5355\u5143 placement \u662f\u82af\u7247\u8bbe\u8ba1\u4e2d\u641c\u7d22\u7a7a\u95f4\u6781\u5927\u3001\u76ee\u6807\u4e92\u76f8\u51b2\u7a81\u7684\u9636\u6bb5\u3002\u4f20\u7edf\u65b9\u6cd5\u4f9d\u8d56\u5de5\u7a0b\u5e08\u53cd\u590d\u8c03 floorplan\u3001\u8dd1\u5b9e\u73b0\u5de5\u5177\u3001\u89c2\u5bdf\u62e5\u585e\u548c\u65f6\u5e8f&#xff0c;\u518d\u4eba\u5de5\u8c03\u6574\u3002\u8be5\u8bba\u6587\u7684\u5173\u952e\u95ee\u9898\u4e0d\u662f\u8ba9\u6a21\u578b\u753b\u7248\u56fe&#xff0c;\u800c\u662f\u628a placement \u53d8\u6210\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u5e8f\u5217\u51b3\u7b56\u95ee\u9898&#xff0c;\u5e76\u8ba9\u7b56\u7565\u80fd\u4ece\u5386\u53f2 block \u8fc1\u79fb\u5230\u65b0 block\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u628a netlist \u8868\u793a\u4e3a\u56fe&#xff0c;\u4f7f\u7528\u56fe\u795e\u7ecf\u7f51\u7edc\u7f16\u7801\u8282\u70b9\u3001\u90bb\u63a5\u5173\u7cfb\u3001\u5b8f\u5355\u5143\u7279\u5f81\u548c\u5de5\u827a\/\u753b\u5e03\u4fe1\u606f\u3002<\/li>\n<li>\u7528\u5f3a\u5316\u5b66\u4e60\u7b56\u7565\u7f51\u7edc\u9010\u4e2a\u653e\u7f6e\u5b8f\u5355\u5143&#xff0c;\u6807\u51c6\u5355\u5143\u518d\u4ea4\u7ed9 force-directed \u65b9\u6cd5\u5904\u7406&#xff0c;\u5956\u52b1\u51fd\u6570\u8fd1\u4f3c wirelength \u4e0e congestion\u3002<\/li>\n<li>\u901a\u8fc7\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u8ba9 policy \u4ece\u4e00\u6279 block \u4e2d\u5b66\u4e60\u53ef\u8fc1\u79fb\u5e03\u5c40\u7ecf\u9a8c&#xff0c;\u800c\u4e0d\u662f\u6bcf\u4e2a block \u90fd\u4ece\u96f6\u5f00\u59cb\u641c\u7d22\u3002<\/li>\n<li>\u628a PPA \u76ee\u6807\u538b\u7f29\u4e3a\u53ef\u4f18\u5316\u4ee3\u7406\u6307\u6807&#xff0c;\u518d\u7528\u5de5\u4e1a EDA \u5de5\u5177\u9a8c\u8bc1\u53ef\u884c\u6027\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u8bba\u6587\u6458\u8981\u660e\u786e\u63d0\u51fa\u5c06 placement \u5efa\u6a21\u4e3a RL \u95ee\u9898&#xff0c;\u5e76\u4ee5 PPA \u4e3a\u4f18\u5316\u76ee\u6807\u3002<\/li>\n<li>\u8bba\u6587\u5185\u8868\u683c\u5bf9\u6bd4 RL\u3001\u6a21\u62df\u9000\u706b\u3001RePlAce \u548c\u4eba\u5de5\u4e13\u5bb6\u65b9\u6848&#xff0c;\u5c55\u793a wirelength\u3001congestion\u3001WNS\u3001\u529f\u8017\u3001\u9762\u79ef\u7b49\u4fe1\u53f7\u3002<\/li>\n<li>\u4f5c\u8005\u58f0\u79f0\u5728\u73b0\u4ee3 accelerator netlist \u4e0a\u53ef\u5728 6 \u5c0f\u65f6\u5185\u751f\u6210\u53ef\u4e0e\u4eba\u5de5\u4e13\u5bb6 comparable \u6216 superhuman \u7684\u5e03\u5c40&#xff0c;\u800c\u4eba\u5de5 baseline \u9700\u8981\u6570\u5468\u4e13\u5bb6\u8fed\u4ee3\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>AlphaChip \u8def\u7ebf\u8bf4\u660e&#xff0c;AI4EDA \u7684\u4e00\u4e2a\u57fa\u7840\u8303\u5f0f\u662f\u201c\u628a\u9ad8\u7ef4\u79bb\u6563\u8bbe\u8ba1\u7a7a\u95f4\u8f6c\u5199\u4e3a\u53ef\u4ea4\u4e92\u7684\u51b3\u7b56\u8fc7\u7a0b\u201d\u3002\u6a21\u578b\u672c\u8eab\u4e0d\u66ff\u4ee3 signoff&#xff0c;\u800c\u662f\u5728\u53ef\u91cf\u5316 reward \u548c\u5de5\u5177\u9a8c\u8bc1\u4e4b\u95f4\u627f\u62c5\u9ad8\u6548\u641c\u7d22\u5668\u89d2\u8272\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a\u82af\u7247 floorplanning \u4ece\u4e13\u5bb6\u7ecf\u9a8c\u9a71\u52a8\u63a8\u8fdb\u5230\u7ecf\u9a8c\u53ef\u8fc1\u79fb\u7684\u7b56\u7565\u5b66\u4e60\u3002 GNN &#043; RL \u7684\u7ec4\u5408\u9002\u914d netlist \u8fd9\u79cd\u5929\u7136\u56fe\u7ed3\u6784\u3002 \u65e9\u671f\u5df2\u7ecf\u628a PPA\u3001\u62e5\u585e\u3001\u65f6\u5e8f\u53ef\u884c\u6027\u7eb3\u5165\u8bba\u6587\u53d9\u4e8b&#xff0c;\u800c\u4e0d\u662f\u53ea\u8ffd\u6c42\u5355\u4e00 wirelength\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u5de5\u7a0b\u4e0a\u6700\u9002\u5408\u7528\u4e8e\u65e9\u671f floorplan \u63a2\u7d22\u3001\u5b8f\u5355\u5143\u521d\u59cb\u5e03\u5c40\u548c\u591a\u65b9\u6848\u751f\u6210\u3002\u5b83\u7684\u4ef7\u503c\u4e0d\u662f\u76f4\u63a5\u7b7e\u6838&#xff0c;\u800c\u662f\u7f29\u77ed\u4e13\u5bb6\u4ece\u7a7a\u767d\u753b\u5e03\u5230\u53ef\u884c\u5e03\u5c40\u5019\u9009\u7684\u65f6\u95f4&#xff0c;\u5e76\u4e3a\u540e\u7eed EDA flow \u63d0\u4f9b\u66f4\u597d\u7684\u8d77\u70b9\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u516c\u5f00\u8bba\u6587\u4e2d\u90e8\u5206\u8bbe\u8ba1\u88ab\u6a21\u7cca\u5904\u7406&#xff0c;\u5546\u4e1a autoplacer baseline\u3001\u5148\u8fdb\u79c1\u6709 TPU block \u548c\u771f\u5b9e signoff \u6761\u4ef6\u96be\u4ee5\u5b8c\u5168\u590d\u73b0\u3002RL reward \u4ecd\u662f\u4ee3\u7406\u76ee\u6807&#xff0c;\u5fc5\u987b\u7531\u5b8c\u6574 STA\u3001routing\u3001IR\/EM\u3001DFM \u6d41\u7a0b\u95ed\u73af\u6821\u9a8c\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5f00\u6e90\u9879\u76ee \/ GPU \u52a0\u901f placement<\/p>\n<h4>DREAMPlace&#xff1a;\u6df1\u5ea6\u5b66\u4e60\u5de5\u5177\u94fe\u542f\u53d1\u7684 GPU \u5206\u6790\u5f0f\u5e03\u5c40<\/h4>\n<p>DREAMPlace GitHub README \u4e0e DAC\/TCAD \u8bba\u6587\u5165\u53e3\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-0430qvs3d0go3.png\" alt=\"DREAMPlace\uff1a\u6df1\u5ea6\u5b66\u4e60\u5de5\u5177\u94fe\u542f\u53d1\u7684 GPU \u5206\u6790\u5f0f\u5e03\u5c40 - DREAMPlace README \u622a\u56fe\" \/><\/p>\n<p>DREAMPlace README \u622a\u56fe \u5b98\u65b9 README \u660e\u786e\u7ed9\u51fa\u9879\u76ee\u5b9a\u4f4d\u3001CPU\/GPU \u652f\u6301\u548c\u52a0\u901f\u58f0\u660e\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u4f20\u7edf placement \u4f18\u5316\u5668\u5728\u5927\u89c4\u6a21\u8bbe\u8ba1\u4e0a\u8fd0\u884c\u65f6\u95f4\u957f&#xff0c;\u4e14\u8bb8\u591a\u6838\u5fc3 kernel \u5177\u6709\u6570\u503c\u8ba1\u7b97\u5bc6\u96c6\u7279\u5f81\u3002DREAMPlace \u5173\u6ce8\u7684\u95ee\u9898\u4e0d\u662f LLM \u751f\u6210\u8bbe\u8ba1&#xff0c;\u800c\u662f\u5982\u4f55\u628a analytical placement \u6620\u5c04\u5230\u73b0\u4ee3 GPU\/\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6&#xff0c;\u4f7f\u4f18\u5316\u5668\u672c\u8eab\u83b7\u5f97\u6570\u91cf\u7ea7\u52a0\u901f\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u628a nonlinear VLSI placement \u7c7b\u6bd4\u4e3a\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u95ee\u9898&#xff0c;\u7528 tensor \u8ba1\u7b97\u548c CUDA kernel \u627f\u8f7d\u5bc6\u5ea6\u3001\u7ebf\u957f\u3001\u7535\u52bf\u7b49\u4f18\u5316\u8fc7\u7a0b\u3002<\/li>\n<li>\u540c\u65f6\u652f\u6301 CPU\/GPU&#xff0c;\u56f4\u7ed5\u5168\u5c40\u5e03\u5c40\u3001legalization\u3001\u8be6\u7ec6\u5e03\u5c40\u6784\u5efa\u53ef\u914d\u7f6e\u5de5\u5177\u94fe\u3002<\/li>\n<li>\u4ee5 PyTorch \u98ce\u683c\u7684\u8f6f\u4ef6\u5de5\u7a0b\u65b9\u5f0f\u66b4\u9732 placement \u8fc7\u7a0b&#xff0c;\u4fbf\u4e8e\u7814\u7a76\u8005\u63d2\u5165\u65b0\u76ee\u6807\u548c\u65b0\u7b97\u5b50\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u9879\u76ee README \u58f0\u79f0\u5728 ISPD 2005 benchmarks \u4e0a&#xff0c;Tesla V100 GPU \u76f8\u5bf9 CPU RePlAce \u5728 global placement\/legalization \u83b7\u5f97 30X \u7ea7\u522b\u52a0\u901f\u3002<\/li>\n<li>README \u540c\u65f6\u8bf4\u660e\u96c6\u6210 GPU \u8be6\u7ec6\u5e03\u5c40\u5668 ABCDPlace&#xff0c;\u5728\u767e\u4e07\u89c4\u6a21 benchmark \u4e0a\u76f8\u5bf9 NTUPlace3 CPU \u7ea6 16X \u52a0\u901f\u3002<\/li>\n<li>\u9879\u76ee\u6301\u7eed\u7ef4\u62a4 timing-driven\u3001multi-electrostatics \u7b49\u540e\u7eed\u65b9\u5411&#xff0c;\u8bf4\u660e\u5176\u4e0d\u662f\u4e00\u6b21\u6027 demo\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>DREAMPlace \u7684\u7406\u8bba\u610f\u4e49\u5728\u4e8e\u8bc1\u660e AI4EDA \u4e0d\u53ea\u6709\u201c\u6a21\u578b\u751f\u6210\u4ee3\u7801\u201d\u4e00\u6761\u8def\u3002\u628a EDA \u4f18\u5316\u76ee\u6807\u6539\u5199\u4e3a\u53ef\u5fae\/\u53ef\u5e76\u884c\u7684 tensor \u7a0b\u5e8f&#xff0c;\u672c\u8d28\u4e0a\u4e5f\u662f AI \u57fa\u7840\u8bbe\u65bd\u5316\u7684\u4e00\u90e8\u5206\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u4ee3\u8868 GPU-first EDA \u4f18\u5316\u5668\u8def\u7ebf\u3002 \u628a\u6df1\u5ea6\u5b66\u4e60\u8f6f\u4ef6\u6808\u7684\u81ea\u52a8\u5e76\u884c\u3001\u5f20\u91cf\u8ba1\u7b97\u548c kernel \u5de5\u7a0b\u8fc1\u79fb\u5230 placement\u3002 \u5bf9\u540e\u7eed RL\/agent placement \u4e5f\u6709\u57fa\u7840\u8bbe\u65bd\u4ef7\u503c&#xff0c;\u56e0\u4e3a\u641c\u7d22\u5916\u5c42\u9700\u8981\u5feb\u901f\u5185\u5c42\u8bc4\u4f30\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u4f5c\u4e3a placement \u7814\u7a76\u7684\u9ad8\u901f\u5b9e\u9a8c\u5e73\u53f0&#xff0c;\u4e5f\u9002\u5408\u4e0e RL\u3001\u8d1d\u53f6\u65af\u4f18\u5316\u3001\u591a\u76ee\u6807\u641c\u7d22\u7ed3\u5408&#xff0c;\u7528\u4e8e\u5feb\u901f\u8bc4\u4f30\u5019\u9009 placement \u53c2\u6570\u6216\u7ea6\u675f\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u5b83\u4e0d\u662f\u81ea\u7136\u8bed\u8a00 agent&#xff0c;\u4e5f\u4e0d\u89e3\u51b3\u89c4\u683c\u7406\u89e3\u548c\u9a8c\u8bc1\u751f\u6210\u95ee\u9898\u3002\u771f\u5b9e\u5546\u4e1a flow \u4e2d\u4ecd\u8981\u5904\u7406 timing closure\u3001routing\u3001power grid\u3001DFM\u3001PDK \u548c signoff \u5de5\u5177\u7ea6\u675f\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5f00\u6e90\u57fa\u7840\u8bbe\u65bd \/ RTL-to-GDS flow<\/p>\n<h4>OpenROAD&#xff1a;\u5f00\u653e RTL-to-GDS \u4e0e no-human-in-loop \u76ee\u6807<\/h4>\n<p>OpenROAD \u5b98\u65b9\u9879\u76ee\u9875\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04jtg5l3pwnoo.png\" alt=\"OpenROAD\uff1a\u5f00\u653e RTL-to-GDS \u4e0e no-human-in-loop \u76ee\u6807 - OpenROAD \u5b98\u65b9\u9875\u9762\u622a\u56fe\" \/><\/p>\n<p>OpenROAD \u5b98\u65b9\u9875\u9762\u622a\u56fe \u5b98\u65b9\u9875\u9762\u5c55\u793a\u5176 democratizing hardware design \u548c 24-hour no-human-in-loop \u76ee\u6807\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>AI4EDA \u7684\u516c\u5f00\u7814\u7a76\u957f\u671f\u53d7\u9650\u4e8e\u5546\u4e1a EDA\/PDK\/license \u4e0d\u53ef\u516c\u5f00\u590d\u73b0\u3002\u6ca1\u6709\u5f00\u653e flow&#xff0c;\u5c31\u5f88\u96be\u6bd4\u8f83 agent\u3001\u6570\u636e\u96c6\u3001\u81ea\u52a8\u8c03\u53c2\u548c\u7aef\u5230\u7aef\u8bbe\u8ba1\u65b9\u6cd5\u3002OpenROAD \u8981\u89e3\u51b3\u7684\u662f\u53ef\u8bbf\u95ee\u3001\u53ef\u590d\u73b0\u3001\u53ef\u81ea\u52a8\u5316\u7684\u8bbe\u8ba1\u5b9e\u73b0\u5e95\u5ea7\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6784\u5efa\u4ece\u7efc\u5408\u3001floorplan\u3001placement\u3001CTS\u3001routing \u5230 signoff \u76f8\u5173\u68c0\u67e5\u7684\u5f00\u653e\u5de5\u5177\u94fe\u3002<\/li>\n<li>\u63d0\u51fa 24 \u5c0f\u65f6\u3001no-human-in-loop layout implementation \u76ee\u6807&#xff0c;\u628a\u5de5\u5177\u81ea\u8c03\u53c2\u3001\u5e76\u884c\u641c\u7d22\u3001\u673a\u5668\u5b66\u4e60\u9884\u6d4b\u4f5c\u4e3a\u7cfb\u7edf\u76ee\u6807\u3002<\/li>\n<li>\u901a\u8fc7\u5f00\u653e\u811a\u672c\u3001\u5f00\u653e\u6570\u636e\u548c\u793e\u533a\u6cbb\u7406\u964d\u4f4e cost\u3001expertise\u3001risk \u4e09\u7c7b\u95e8\u69db\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u5b98\u65b9\u9875\u9762\u5c06\u76ee\u6807\u8868\u8ff0\u4e3a no-human-in-loop\u300124-hour layout design&#xff0c;\u5e76\u5f3a\u8c03\u65e0 PPA loss \u7684 tapeout-capable tools\u3002<\/li>\n<li>OpenROAD \u4e5f\u662f EDA Corpus\u3001ORAssistant\u3001\u8bb8\u591a LLM4EDA \u811a\u672c\u751f\u6210\u5de5\u4f5c\u7684\u5e95\u5c42\u5b9e\u9a8c\u73af\u5883\u3002<\/li>\n<li>\u5176\u4ef7\u503c\u4e0d\u53ea\u5728\u5de5\u5177\u672c\u8eab&#xff0c;\u800c\u5728\u5f62\u6210\u53ef\u516c\u5f00\u53d1\u5e03\u6570\u636e\u3001\u4efb\u52a1\u548c\u53ef\u590d\u73b0\u5b9e\u9a8c\u7684\u516c\u5171\u57fa\u7840\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>OpenROAD \u8bf4\u660e AI4EDA \u7684\u53ef\u590d\u73b0\u6027\u6765\u81ea\u5de5\u5177\u94fe\u5f00\u653e&#xff0c;\u800c\u4e0d\u4ec5\u6765\u81ea\u6a21\u578b\u5f00\u653e\u3002\u6ca1\u6709\u53ef\u6267\u884c flow&#xff0c;LLM \u53ea\u80fd\u505c\u7559\u5728\u6587\u672c\u751f\u6210&#xff1b;\u6709\u4e86\u5f00\u653e flow&#xff0c;agent \u624d\u80fd\u88ab\u771f\u5b9e\u5de5\u5177\u53cd\u9988\u7ea6\u675f\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a EDA \u4ece\u5c01\u95ed\u8f6f\u4ef6\u6808\u7684\u4e00\u6b21\u6027\u7ecf\u9a8c\u8f6c\u6210\u53ef\u5171\u4eab\u7684\u5b9e\u9a8c\u5e73\u53f0\u3002 \u8ba9\u516c\u5f00 LLM \u6570\u636e\u96c6\u80fd\u591f\u7ed1\u5b9a\u771f\u5b9e\u5de5\u5177 API \u548c\u811a\u672c\u6267\u884c\u3002 \u5bf9\u9ad8\u6821\u3001\u521d\u521b\u548c\u5f00\u6e90\u786c\u4ef6\u751f\u6001\u5177\u6709\u57fa\u7840\u8bbe\u65bd\u610f\u4e49\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u642d\u5efa\u5185\u90e8 AI4EDA prototype&#xff1a;RAG \u95ee\u7b54\u3001\u811a\u672c\u751f\u6210\u3001flow \u53c2\u6570\u641c\u7d22\u3001\u9519\u8bef\u65e5\u5fd7\u8bca\u65ad\u548c\u5c0f\u89c4\u6a21 PPA \u5b9e\u9a8c\u90fd\u53ef\u4ee5\u5148\u5728 OpenROAD \u4e0a\u95ed\u73af\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u5f00\u653e flow \u4e0e\u5148\u8fdb\u5546\u4e1a\u8282\u70b9\u4e4b\u95f4\u4ecd\u6709 PDK\u3001rule deck\u3001QoR\u3001IP\u3001signoff \u7cbe\u5ea6\u5dee\u8ddd\u3002\u5bf9\u9ad8\u7aef SoC&#xff0c;OpenROAD \u66f4\u9002\u5408\u4f5c\u4e3a\u7814\u7a76\u5e73\u53f0\u548c\u65b9\u6cd5\u9a8c\u8bc1\u5e95\u5ea7&#xff0c;\u800c\u4e0d\u662f\u5b8c\u6574\u66ff\u4ee3\u5546\u4e1a\u5b9e\u73b0 flow\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5b66\u672f\u8bba\u6587 &#043; \u5f00\u6e90\u6570\u636e \/ LLM4EDA \u6570\u636e\u5de5\u7a0b<\/p>\n<h4>EDA Corpus&#xff1a;\u9762\u5411 OpenROAD \u4ea4\u4e92\u7684 LLM \u6570\u636e\u96c6<\/h4>\n<p>arXiv 2405.06676 \u4e0e OpenROAD-Assistant\/EDA-Corpus GitHub\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04qiv5lgecprc.png\" alt=\"EDA Corpus\uff1a\u9762\u5411 OpenROAD \u4ea4\u4e92\u7684 LLM \u6570\u636e\u96c6 - prompt-script \u6837\u4f8b\u622a\u56fe\" \/><\/p>\n<p>prompt-script \u6837\u4f8b\u622a\u56fe \u00b7 PDF p.2 \u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a\u81ea\u7136\u8bed\u8a00 IR drop \u4efb\u52a1\u5982\u4f55\u6620\u5c04\u4e3a OpenROAD Python \u811a\u672c\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04qrc3ustduy1.png\" alt=\"EDA Corpus\uff1a\u9762\u5411 OpenROAD \u4ea4\u4e92\u7684 LLM \u6570\u636e\u96c6 - \u6570\u636e\u7edf\u8ba1\u8868\u622a\u56fe\" \/><\/p>\n<p>\u6570\u636e\u7edf\u8ba1\u8868\u622a\u56fe \u00b7 PDF p.3 \u8bba\u6587\u5185\u90e8\u8868\u683c\u7ed9\u51fa question-answer \u4e0e prompt-script \u6570\u636e\u7c7b\u522b\u548c\u89c4\u6a21\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-042ttthzpnpvo.png\" alt=\"EDA Corpus\uff1a\u9762\u5411 OpenROAD \u4ea4\u4e92\u7684 LLM \u6570\u636e\u96c6 - GitHub README \u622a\u56fe\" \/><\/p>\n<p>GitHub README \u622a\u56fe README \u8bf4\u660e\u6570\u636e\u96c6\u7c7b\u578b\u3001\u589e\u5f3a\u89c4\u6a21\u548c\u8bb8\u53ef\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u901a\u7528 LLM \u5373\u4f7f\u4f1a\u5199 Python&#xff0c;\u4e5f\u672a\u5fc5\u77e5\u9053 EDA \u5de5\u5177\u547d\u4ee4\u3001OpenROAD \u5bf9\u8c61\u6a21\u578b\u3001\u7269\u7406\u8bbe\u8ba1\u672f\u8bed\u548c\u5e38\u89c1 flow \u4efb\u52a1\u3002EDA Corpus \u89e3\u51b3\u7684\u662f\u201c\u6a21\u578b\u5982\u4f55\u83b7\u5f97\u53ef\u8bb8\u53ef\u3001\u53ef\u8bad\u7ec3\u3001\u53ef\u6267\u884c\u7684 EDA \u4ea4\u4e92\u6570\u636e\u201d\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u56f4\u7ed5 OpenROAD \u6784\u5efa\u4e24\u7c7b\u6837\u672c&#xff1a;question-answer \u548c prompt-script\u3002<\/li>\n<li>\u4ece\u6587\u6863\u3001issue\u3001\u8ba8\u8bba\u3001\u5de5\u5177\u4f7f\u7528\u573a\u666f\u4e2d\u6574\u7406\u95ee\u7b54&#xff0c;\u5e76\u628a\u81ea\u7136\u8bed\u8a00\u4efb\u52a1\u6620\u5c04\u4e3a OpenROAD Python \u811a\u672c\u3002<\/li>\n<li>\u901a\u8fc7\u6539\u5199 prompt\u3001\u6539\u53d8\u91cf\u540d\/\u53c2\u6570\u7b49\u65b9\u5f0f\u505a\u6570\u636e\u589e\u5f3a&#xff0c;\u63d0\u9ad8\u8bed\u8a00\u591a\u6837\u6027\u548c\u811a\u672c\u8986\u76d6\u9762\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u8bba\u6587\u6458\u8981\u8bf4\u660e\u6570\u636e\u96c6\u8d85\u8fc7 1000 \u4e2a datapoints&#xff0c;\u5305\u542b\u95ee\u7b54\u548c prompt-script \u4e24\u79cd\u683c\u5f0f\u3002<\/li>\n<li>GitHub README \u7ed9\u51fa\u975e\u589e\u5f3a\/\u589e\u5f3a\u6570\u636e\u89c4\u6a21&#xff1a;QA 198\/590\u3001PS 395\/943\u3001Combined 593\/1533\u3002<\/li>\n<li>\u8bba\u6587\u5185\u622a\u56fe\u5c55\u793a IR drop analysis \u7684 prompt-script \u6837\u4f8b\u548c\u6570\u636e\u7edf\u8ba1\u8868&#xff0c;\u8bf4\u660e\u6570\u636e\u4e0d\u662f\u62bd\u8c61\u95ee\u7b54&#xff0c;\u800c\u662f\u7ed1\u5b9a\u5de5\u5177\u64cd\u4f5c\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>EDA Corpus \u7684\u7ed3\u8bba\u662f&#xff0c;EDA \u9886\u57df LLM \u7684\u6838\u5fc3\u8d44\u4ea7\u4e0d\u662f\u6cdb\u5316\u8bed\u6599&#xff0c;\u800c\u662f\u201c\u5de5\u5177\u52a8\u4f5c\u8bed\u6599\u201d\u3002\u53ea\u6709\u628a\u81ea\u7136\u8bed\u8a00\u3001\u5de5\u5177 API\u3001\u811a\u672c\u548c\u6267\u884c\u8bed\u4e49\u5bf9\u9f50&#xff0c;agent \u624d\u53ef\u80fd\u7a33\u5b9a\u8fdb\u5165\u771f\u5b9e flow\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a OpenROAD \u4ece\u5de5\u5177\u53d8\u6210 LLM \u53ef\u5b66\u4e60\u7684\u4ea4\u4e92\u73af\u5883\u3002 \u5f3a\u8c03\u8bb8\u53ef\u53ef\u7528\u548c\u516c\u5f00\u53d1\u5e03&#xff0c;\u964d\u4f4e\u9886\u57df\u6a21\u578b\u590d\u73b0\u5b9e\u9a8c\u95e8\u69db\u3002 \u5bf9\u4f01\u4e1a\u5185\u90e8\u5efa\u8bbe EDA copilot \u6570\u636e\u7ba1\u7ebf\u6709\u76f4\u63a5\u53c2\u8003\u4ef7\u503c\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u53ef\u4f5c\u4e3a\u4f01\u4e1a\u6784\u5efa EDA RAG\/agent \u6570\u636e\u96c6\u7684\u6a21\u677f&#xff1a;\u6309\u5de5\u5177\u547d\u4ee4\u3001\u8bbe\u8ba1\u9636\u6bb5\u3001\u9519\u8bef\u7c7b\u578b\u3001\u811a\u672c\u7247\u6bb5\u3001\u9a8c\u8bc1\u7ed3\u679c\u6765\u7ec4\u7ec7\u8bed\u6599&#xff0c;\u800c\u4e0d\u662f\u7b80\u5355\u6536\u96c6 PDF \u6216 wiki\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u6570\u636e\u5f3a\u7ed1\u5b9a OpenROAD&#xff0c;\u8fc1\u79fb\u5230\u5546\u4e1a\u5de5\u5177\u9700\u8981\u91cd\u5efa API\u3001license\u3001\u65e5\u5fd7\u3001\u9519\u8bef\u4fe1\u606f\u548c\u5185\u90e8\u89c4\u8303\u6620\u5c04\u3002\u6570\u636e\u589e\u5f3a\u4e0d\u80fd\u66ff\u4ee3\u771f\u5b9e\u5931\u8d25\u8f68\u8ff9\u548c\u771f\u5b9e\u8bbe\u8ba1\u7ea6\u675f\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5b66\u672f\u8bba\u6587 &#043; \u9879\u76ee\u9875 \/ \u7269\u7406\u8bbe\u8ba1\u6570\u636e\u96c6<\/p>\n<h4>CircuitNet&#xff1a;\u9762\u5411\u7269\u7406\u8bbe\u8ba1\u9884\u6d4b\u4efb\u52a1\u7684\u5f00\u653e\u6570\u636e\u96c6<\/h4>\n<p>arXiv 2208.01040 \u4e0e CircuitNet \u9879\u76ee\u9875\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04gqeocszbk2j.png\" alt=\"CircuitNet\uff1a\u9762\u5411\u7269\u7406\u8bbe\u8ba1\u9884\u6d4b\u4efb\u52a1\u7684\u5f00\u653e\u6570\u636e\u96c6 - \u6570\u636e\u4e0e\u4efb\u52a1\u603b\u89c8\u56fe\u622a\u56fe\" \/><\/p>\n<p>\u6570\u636e\u4e0e\u4efb\u52a1\u603b\u89c8\u56fe\u622a\u56fe \u00b7 PDF p.2 \u8bba\u6587\u5185\u90e8\u56fe\u628a\u6570\u636e\u7edf\u8ba1\u3001\u7279\u5f81\u62bd\u53d6\u9636\u6bb5\u548c\u9884\u6d4b\u4efb\u52a1\u653e\u5728\u540c\u4e00\u6846\u67b6\u4e2d\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-044jpgl5z24pn.png\" alt=\"CircuitNet\uff1a\u9762\u5411\u7269\u7406\u8bbe\u8ba1\u9884\u6d4b\u4efb\u52a1\u7684\u5f00\u653e\u6570\u636e\u96c6 - \u8bbe\u8ba1\u7edf\u8ba1\u8868\u622a\u56fe\" \/><\/p>\n<p>\u8bbe\u8ba1\u7edf\u8ba1\u8868\u622a\u56fe \u00b7 PDF p.2 \u8bba\u6587\u5185\u90e8\u8868\u683c\u5c55\u793a netlist\u3001\u7efc\u5408\u53d8\u5316\u548c\u7269\u7406\u8bbe\u8ba1\u53d8\u5316\u7b49\u6570\u636e\u7ef4\u5ea6\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04sjcddcymxlp.png\" alt=\"CircuitNet\uff1a\u9762\u5411\u7269\u7406\u8bbe\u8ba1\u9884\u6d4b\u4efb\u52a1\u7684\u5f00\u653e\u6570\u636e\u96c6 - CircuitNet \u9879\u76ee\u9875\u622a\u56fe\" \/><\/p>\n<p>CircuitNet \u9879\u76ee\u9875\u622a\u56fe \u9879\u76ee\u9875\u4f5c\u4e3a\u6570\u636e\u53d1\u5e03\u548c\u4efb\u52a1\u5165\u53e3\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u7269\u7406\u8bbe\u8ba1\u4e2d\u7684\u62e5\u585e\u3001DRC\u3001IR drop\u3001\u529f\u8017\u7b49\u9884\u6d4b\u4efb\u52a1\u9700\u8981\u5927\u91cf\u7248\u56fe\u548c\u4e2d\u95f4\u7279\u5f81&#xff0c;\u4f46\u516c\u5f00\u6570\u636e\u7a00\u7f3a\u3002CircuitNet \u89e3\u51b3\u7684\u662f\u5982\u4f55\u628a RTL-to-layout flow \u4e2d\u7684\u591a\u9636\u6bb5\u7279\u5f81\u8f6c\u6210\u53ef\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u4ece\u8bbe\u8ba1\u6d41\u7a0b\u4e2d\u62bd\u53d6 netlist\u3001placement\u3001routing \u524d\u540e\u7279\u5f81\u3001\u56fe\u50cf\u5316 map \u548c\u6807\u7b7e\u3002<\/li>\n<li>\u56f4\u7ed5\u8de8\u9636\u6bb5\u9884\u6d4b\u6784\u5efa benchmark&#xff0c;\u4f7f\u6a21\u578b\u80fd\u5728\u65e9\u671f\u9636\u6bb5\u9884\u6d4b\u540e\u671f\u98ce\u9669\u3002<\/li>\n<li>\u901a\u8fc7\u9879\u76ee\u9875\u53d1\u5e03\u6570\u636e\u548c\u4efb\u52a1&#xff0c;\u4f7f\u62e5\u585e\u9884\u6d4b\u3001DRC \u9884\u6d4b\u3001IR drop \u9884\u6d4b\u7b49\u7814\u7a76\u53ef\u6bd4\u8f83\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>arXiv \u6458\u8981\u79f0\u5176\u4e3a VLSI CAD \u673a\u5668\u5b66\u4e60\u4efb\u52a1\u7684\u9996\u4e2a open-source dataset\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a\u8bbe\u8ba1\u7edf\u8ba1\u3001\u7279\u5f81\u62bd\u53d6\u9636\u6bb5\u548c\u9884\u6d4b\u4efb\u52a1\u7684\u5173\u7cfb\u3002<\/li>\n<li>\u9879\u76ee\u9875\u63d0\u4f9b\u6570\u636e\u5165\u53e3\u548c\u4efb\u52a1\u8bf4\u660e&#xff0c;\u8bf4\u660e\u5176\u9762\u5411\u957f\u671f benchmark \u4f7f\u7528\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>CircuitNet \u7684\u6838\u5fc3\u7ed3\u8bba\u662f&#xff1a;AI4EDA \u540e\u7aef\u80fd\u529b\u4e0d\u662f\u4ece\u81ea\u7136\u8bed\u8a00\u91cc\u957f\u51fa\u6765\u7684&#xff0c;\u800c\u662f\u4ece flow trace\u3001layout map\u3001netlist graph \u548c PPA label \u4e2d\u5b66\u51fa\u6765\u7684\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a\u7269\u7406\u8bbe\u8ba1\u4ece\u4e0d\u53ef\u89c1\u7684\u5546\u4e1a\u6d41\u7a0b\u8f6c\u6210\u516c\u5f00\u6570\u636e\u4efb\u52a1\u3002 \u7a81\u51fa\u8de8\u9636\u6bb5\u9884\u6d4b&#xff0c;\u8ba9\u65e9\u671f\u8bbe\u8ba1\u51b3\u7b56\u80fd\u770b\u5230\u540e\u671f\u98ce\u9669\u3002 \u4e3a\u591a\u6a21\u6001 PD assistant \u548c layout-aware LLM \u63d0\u4f9b\u8bad\u7ec3\u524d\u63d0\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u7528\u4e8e\u8bad\u7ec3\u62e5\u585e\/DRC\/IR \u98ce\u9669\u9884\u6d4b\u5668&#xff0c;\u4e5f\u9002\u5408\u8bc4\u4f30\u89c6\u89c9\u6a21\u578b\u3001GNN \u548c\u591a\u6a21\u6001 agent \u662f\u5426\u7406\u89e3\u7269\u7406\u8bbe\u8ba1\u8bc1\u636e\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u516c\u5f00\u6570\u636e\u89c4\u6a21\u548c\u8bbe\u8ba1\u5206\u5e03\u4e0d\u80fd\u8986\u76d6\u5148\u8fdb\u5546\u4e1a SoC \u7684\u5168\u90e8\u590d\u6742\u6027\u3002\u6a21\u578b\u5728 CircuitNet \u4e0a\u6709\u6548&#xff0c;\u4e0d\u7b49\u4e8e\u5728\u79c1\u6709 PDK\u3001\u771f\u5b9e IP\u3001\u590d\u6742\u7ea6\u675f\u548c signoff corner \u4e0b\u76f4\u63a5\u6709\u6548\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5b66\u672f\u8bba\u6587 \/ \u591a\u6a21\u6001 EDA \u6570\u636e\u96c6<\/p>\n<h4>ForgeEDA&#xff1a;\u591a\u6a21\u6001 EDA \u6570\u636e\u96c6\u4e0e\u8868\u793a\u7edf\u4e00<\/h4>\n<p>arXiv 2505.02016\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-0421rjzi5gdyy.png\" alt=\"ForgeEDA\uff1a\u591a\u6a21\u6001 EDA \u6570\u636e\u96c6\u4e0e\u8868\u793a\u7edf\u4e00 - ForgeEDA \u603b\u89c8\u56fe\u622a\u56fe\" \/><\/p>\n<p>ForgeEDA \u603b\u89c8\u56fe\u622a\u56fe \u00b7 PDF p.2 \u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a\u591a\u6a21\u6001\u6570\u636e\u4ece\u4ee3\u7801\u4ed3\u5e93\u5230\u903b\u8f91\/\u7269\u7406\u8868\u793a\u7684\u7ec4\u7ec7\u65b9\u5f0f\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04qqaqu1zl4c5.png\" alt=\"ForgeEDA\uff1a\u591a\u6a21\u6001 EDA \u6570\u636e\u96c6\u4e0e\u8868\u793a\u7edf\u4e00 - \u6570\u636e\u5185\u5bb9\u8868\u622a\u56fe\" \/><\/p>\n<p>\u6570\u636e\u5185\u5bb9\u8868\u622a\u56fe \u00b7 PDF p.4 \u8bba\u6587\u5185\u90e8\u8868\u683c\u5c55\u793a collected designs \u7684\u7c7b\u522b\u3001\u6837\u672c\u91cf\u548c\u6a21\u6001\u7ec4\u6210\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u8bb8\u591a AI4EDA \u4efb\u52a1\u5404\u81ea\u4f7f\u7528 RTL\u3001AIG\u3001netlist\u3001placement\u3001timing\/PPA report&#xff0c;\u4f46\u7f3a\u5c11\u8de8\u8868\u793a\u7edf\u4e00\u6570\u636e\u3002ForgeEDA \u8bd5\u56fe\u89e3\u51b3\u201c\u540c\u4e00\u4e2a\u8bbe\u8ba1\u5982\u4f55\u8de8\u903b\u8f91\u3001\u6620\u5c04\u3001\u653e\u7f6e\u548c\u62a5\u544a\u5f62\u6210\u591a\u6a21\u6001\u8bad\u7ec3\u6837\u672c\u201d\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6536\u96c6 RTL code\u3001post-mapping netlist\u3001AIG\u3001placed netlist \u7b49\u591a\u79cd\u8868\u793a\u3002<\/li>\n<li>\u56f4\u7ed5 PPA \u4f18\u5316\u3001AIG optimization\u3001technology mapping \u7b49\u4efb\u52a1\u6784\u5efa benchmark\u3002<\/li>\n<li>\u7528\u540c\u4e00\u8bbe\u8ba1\u7684\u591a\u9636\u6bb5\u8868\u793a\u5e2e\u52a9\u6a21\u578b\u5b66\u4e60\u8de8\u9636\u6bb5\u8bed\u4e49&#xff0c;\u800c\u975e\u53ea\u5b66\u4e60\u5355\u4e00\u6587\u672c\u6216\u56fe\u50cf\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u8bba\u6587\u6458\u8981\u5217\u51fa RTL\u3001PM netlists\u3001AIGs\u3001placed netlists \u7b49\u8868\u793a&#xff0c;\u5e76\u5f3a\u8c03\u7528\u4e8e PPA optimization\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8 overview \u56fe\u5c55\u793a\u4ece RTL code repository \u5230 AIG\u3001timing report\u3001placed netlist\u3001sys report \u7684\u6570\u636e\u7ec4\u7ec7\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u8868\u683c\u7ed9\u51fa collected designs \u548c\u4e0d\u540c circuit collection \u7684\u6570\u91cf\/\u6a21\u6001\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>ForgeEDA \u8868\u660e\u4e0b\u4e00\u9636\u6bb5 EDA foundation model \u7684\u8f93\u5165\u4e0d\u4f1a\u662f\u5355\u4e00 HDL \u6587\u672c&#xff0c;\u800c\u4f1a\u662f\u8de8\u5c42\u7ea7\u3001\u8de8\u6a21\u6001\u3001\u8de8\u5de5\u5177\u9636\u6bb5\u7684\u7edf\u4e00\u8bbe\u8ba1\u8868\u793a\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a\u201c\u8bbe\u8ba1\u5bf9\u8c61\u201d\u4ece\u4e00\u6bb5\u4ee3\u7801\u6269\u5c55\u4e3a\u591a\u9636\u6bb5 representation bundle\u3002 \u8ba9 PPA \u4f18\u5316\u548c\u903b\u8f91\u4f18\u5316\u53ef\u4ee5\u5171\u4eab\u7edf\u4e00\u6570\u636e\u5165\u53e3\u3002 \u5bf9\u6784\u5efa\u8bbe\u8ba1\u5411\u91cf\u5e93\u3001\u68c0\u7d22\u76f8\u4f3c block \u548c\u8de8\u9636\u6bb5\u98ce\u9669\u9884\u6d4b\u6709\u542f\u53d1\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u4f01\u4e1a\u53ef\u4ee5\u501f\u9274\u5176 schema&#xff1a;\u5bf9\u6bcf\u4e2a block \u56fa\u5b9a\u4fdd\u5b58 RTL\u3001netlist\u3001AIG\u3001placement\u3001timing\u3001power\u3001\u811a\u672c\u548c\u5de5\u5177\u7248\u672c&#xff0c;\u4f5c\u4e3a\u5185\u90e8 AI4EDA \u6570\u636e\u6e56\u7684\u57fa\u672c\u884c\u683c\u5f0f\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u516c\u5f00\u6570\u636e\u4ecd\u96be\u8986\u76d6\u5546\u4e1a IP\u3001\u5148\u8fdb\u8282\u70b9 rule \u548c\u771f\u5b9e\u5de5\u7a0b\u53d8\u66f4\u5386\u53f2\u3002\u591a\u6a21\u6001\u6570\u636e\u8d8a\u590d\u6742&#xff0c;\u8d8a\u9700\u8981\u4e25\u683c\u7248\u672c\u7ba1\u7406\u3001\u5de5\u5177 provenance \u548c\u6807\u7b7e\u4e00\u81f4\u6027\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5b66\u672f\u8bba\u6587 \/ LLM4RTL \u65e9\u671f\u6848\u4f8b<\/p>\n<h4>Chip-Chat&#xff1a;\u5bf9\u8bdd\u5f0f\u786c\u4ef6\u8bbe\u8ba1\u4e0e AI-written HDL tapeout \u6848\u4f8b<\/h4>\n<p>arXiv 2305.13243\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04vzqtdpqlt1r.png\" alt=\"Chip-Chat\uff1a\u5bf9\u8bdd\u5f0f\u786c\u4ef6\u8bbe\u8ba1\u4e0e AI-written HDL tapeout \u6848\u4f8b - \u8bbe\u8ba1 prompt \u622a\u56fe\" \/><\/p>\n<p>\u8bbe\u8ba1 prompt \u622a\u56fe \u00b7 PDF p.2 \u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a 8-bit processor \u7684\u8d77\u59cb\u8bbe\u8ba1 prompt \u548c\u5bf9\u8bdd\u5165\u53e3\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04je14cibmzpk.png\" alt=\"Chip-Chat\uff1a\u5bf9\u8bdd\u5f0f\u786c\u4ef6\u8bbe\u8ba1\u4e0e AI-written HDL tapeout \u6848\u4f8b - ISA\/\u5bf9\u8bdd\u8868\u622a\u56fe\" \/><\/p>\n<p>ISA\/\u5bf9\u8bdd\u8868\u622a\u56fe \u00b7 PDF p.3 \u8bba\u6587\u5185\u90e8\u8868\u683c\u548c\u4ee3\u7801\u622a\u56fe\u8bf4\u660e\u8be5\u6848\u4f8b\u4f9d\u8d56\u591a\u8f6e\u4ea4\u4e92\u4e0e\u4eba\u5de5\u5f15\u5bfc\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u81ea\u7136\u8bed\u8a00\u89c4\u683c\u5230 HDL \u7684\u5173\u952e\u96be\u70b9\u5728\u4e8e\u89c4\u683c\u4e0d\u5b8c\u6574\u3001\u6a21\u5757\u63a5\u53e3\u4f1a\u6f02\u79fb\u3001bug \u4fee\u590d\u9700\u8981\u4e0a\u4e0b\u6587\u8bb0\u5fc6\u3002Chip-Chat \u7814\u7a76\u7684\u95ee\u9898\u662f&#xff1a;\u5546\u4e1a\u5bf9\u8bdd LLM \u662f\u5426\u80fd\u5728\u5de5\u7a0b\u5e08\u4ea4\u4e92\u4e0b\u5b8c\u6210\u4e00\u4e2a\u5c0f\u578b\u5904\u7406\u5668\u7684 HDL \u8bbe\u8ba1\u5e76\u8fdb\u5165 tapeout\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u5de5\u7a0b\u5e08\u901a\u8fc7\u591a\u8f6e\u5bf9\u8bdd\u4e0e LLM \u534f\u540c\u5b9a\u4e49 8-bit accumulator-based microprocessor\u3002<\/li>\n<li>LLM \u751f\u6210 Verilog \u6a21\u5757&#xff0c;\u5de5\u7a0b\u5e08\u901a\u8fc7\u540e\u7eed prompt \u5f15\u5bfc\u63a5\u53e3\u4fee\u6b63\u3001bug \u4fee\u590d\u548c\u7ed3\u6784\u8c03\u6574\u3002<\/li>\n<li>\u6700\u7ec8\u628a\u751f\u6210 HDL \u9001\u5165 SkyWater 130nm shuttle tapeout&#xff0c;\u5f62\u6210\u5f3a\u6848\u4f8b\u4fe1\u53f7\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>arXiv \u6458\u8981\u79f0\u8be5\u6848\u4f8b\u662f\u4f5c\u8005\u8ba4\u4e3a\u7684 world \u2019 s first wholly-AI-written HDL for tapeout\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a\u8d77\u59cb\u8bbe\u8ba1 prompt\u3001bug-fix conversation\u3001\u6570\u636e\u901a\u8def\u8bf4\u660e\u548c\u7efc\u5408\u4fe1\u606f\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u8868\u683c\/\u622a\u56fe\u5c55\u793a ISA\u3001\u5bf9\u8bdd\u8fc7\u7a0b\u548c\u4ee3\u7801\u7247\u6bb5&#xff0c;\u53cd\u6620\u8be5\u6d41\u7a0b\u9ad8\u5ea6\u4f9d\u8d56\u4ea4\u4e92\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>Chip-Chat \u7684\u7ed3\u8bba\u4e0d\u662f\u201cLLM \u5df2\u80fd\u72ec\u7acb\u5b8c\u6210\u82af\u7247\u8bbe\u8ba1\u201d&#xff0c;\u800c\u662f\u201cLLM \u53ef\u4ee5\u6210\u4e3a\u786c\u4ef6\u5de5\u7a0b\u5e08\u7684\u4ea4\u4e92\u5f0f\u8349\u56fe\u548c\u4fee\u590d\u4f19\u4f34\u201d\u3002\u6210\u529f\u6761\u4ef6\u662f\u4efb\u52a1\u5c0f\u3001\u53cd\u9988\u5bc6\u3001\u5de5\u7a0b\u5e08\u5f3a\u76d1\u7763\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a LLM4RTL \u4ece benchmark \u63a8\u5230 tapeout \u53d9\u4e8b\u5c42\u9762\u3002 \u63ed\u793a prompt\u3001\u4e0a\u4e0b\u6587\u4fdd\u6301\u3001\u63a5\u53e3\u4e00\u81f4\u6027\u548c bug \u4fee\u590d\u662f\u6838\u5fc3\u75db\u70b9\u3002 \u4e3a\u540e\u7eed RTLFixer\u3001VerilogCoder\u3001ACE-RTL \u7b49\u95ed\u73af agent \u94fa\u57ab\u4e86\u95ee\u9898\u5b9a\u4e49\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u7528\u4e8e\u6559\u5b66\u3001\u5c0f\u578b IP \u539f\u578b\u3001\u521d\u59cb RTL \u8349\u7a3f\u548c\u9700\u6c42\u6f84\u6e05&#xff0c;\u4e0d\u9002\u5408\u4f5c\u4e3a\u65e0\u4eba\u503c\u5b88 signoff \u8bbe\u8ba1\u6d41\u7a0b\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u6848\u4f8b\u89c4\u6a21\u5f88\u5c0f&#xff0c;\u4e14\u5de5\u7a0b\u5e08\u5728\u56de\u8def\u4e2d\u505a\u4e86\u5927\u91cf prompt \u5206\u89e3\u548c\u4fee\u6b63\u3002\u5176 tapeout \u4fe1\u53f7\u5177\u6709\u8c61\u5f81\u610f\u4e49&#xff0c;\u4f46\u4e0d\u80fd\u5916\u63a8\u5230\u590d\u6742 SoC\u3001\u4f4e\u529f\u8017\u3001CDC\/RDC\u3001DFT \u548c\u5b89\u5168\u5173\u952e\u8bbe\u8ba1\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5b66\u672f\u8bba\u6587 \/ NLP-to-RTL \u6846\u67b6<\/p>\n<h4>ChipGPT&#xff1a;\u81ea\u7136\u8bed\u8a00\u786c\u4ef6\u8bbe\u8ba1\u7684\u56db\u9636\u6bb5 zero-code \u6846\u67b6<\/h4>\n<p>arXiv 2305.14019\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-045a03rdjq54f.png\" alt=\"ChipGPT\uff1a\u81ea\u7136\u8bed\u8a00\u786c\u4ef6\u8bbe\u8ba1\u7684\u56db\u9636\u6bb5 zero-code \u6846\u67b6 - ChipGPT \u6846\u67b6\u56fe\u622a\u56fe\" \/><\/p>\n<p>ChipGPT \u6846\u67b6\u56fe\u622a\u56fe \u00b7 PDF p.3 \u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a four-stage zero-code logic design framework\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04eng2egqpmpx.png\" alt=\"ChipGPT\uff1a\u81ea\u7136\u8bed\u8a00\u786c\u4ef6\u8bbe\u8ba1\u7684\u56db\u9636\u6bb5 zero-code \u6846\u67b6 - \u5b9e\u9a8c\u7ed3\u679c\u8868\u622a\u56fe\" \/><\/p>\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u622a\u56fe \u00b7 PDF p.6 \u8bba\u6587\u5185\u90e8\u8868\u683c\u5c55\u793a\u4e0d\u540c\u5019\u9009\u8bbe\u8ba1\u5728\u9762\u79ef\u3001\u529f\u8017\u3001\u6027\u80fd\u7b49\u6307\u6807\u4e0b\u7684\u5dee\u5f02\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u76f4\u63a5\u8ba9 LLM \u4ece\u81ea\u7136\u8bed\u8a00\u751f\u6210 RTL&#xff0c;\u5e38\u89c1\u95ee\u9898\u662f\u63a5\u53e3\u4e0d\u7a33\u3001\u8bed\u6cd5\u9519\u8bef\u3001\u529f\u80fd\u4e0d\u5b8c\u6574\u3001\u5019\u9009\u7a7a\u95f4\u4e0d\u53ef\u63a7\u3002ChipGPT \u7814\u7a76\u7684\u95ee\u9898\u662f\u5982\u4f55\u901a\u8fc7\u6846\u67b6\u5316\u524d\u5904\u7406\u3001\u8f93\u51fa\u7ba1\u7406\u548c\u641c\u7d22&#xff0c;\u8ba9 zero-code \u903b\u8f91\u8bbe\u8ba1\u66f4\u53ef\u63a7\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>Specification Split \u548c Prompt Manager \u8d1f\u8d23\u628a\u81ea\u7136\u8bed\u8a00\u9700\u6c42\u62c6\u6210\u66f4\u9002\u5408 LLM \u7684 prompt\u3002<\/li>\n<li>LLM \u751f\u6210\u521d\u59cb Verilog&#xff0c;Output Manager \u505a\u6821\u6b63\u548c\u4f18\u5316\u3002<\/li>\n<li>Enumerative Search \u5728\u5019\u9009\u8bbe\u8ba1\u7a7a\u95f4\u4e2d\u641c\u7d22\u6ee1\u8db3\u76ee\u6807\u6307\u6807\u7684\u8bbe\u8ba1\u3002<\/li>\n<li>\u7528 programmability\u3001controllability\u3001design space \u7b49\u7ef4\u5ea6\u8bc4\u4f30 NL-to-RTL \u6548\u679c\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>arXiv \u6458\u8981\u79f0 ChipGPT \u662f scalable four-stage zero-code logic design framework\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u6846\u67b6\u56fe\u5c55\u793a split\u3001prompt manager\u3001output manager \u548c search \u7684\u534f\u4f5c\u5173\u7cfb\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u8868\u683c\u5c55\u793a\u4e0d\u540c\u8bbe\u8ba1\u548c prompt\/search \u914d\u7f6e\u4e0b\u7684\u7ed3\u679c\u5dee\u5f02\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>ChipGPT \u7684\u5173\u952e\u7ed3\u8bba\u662f&#xff0c;LLM \u751f\u6210\u786c\u4ef6\u4e0d\u80fd\u53ea\u4f9d\u8d56\u5355\u6b21 prompt\u3002\u5fc5\u987b\u628a\u81ea\u7136\u8bed\u8a00\u89c4\u683c\u53d8\u6210\u53ef\u7ba1\u7406\u7684\u641c\u7d22\u95ee\u9898&#xff0c;\u5e76\u7528\u540e\u5904\u7406\u548c\u76ee\u6807\u51fd\u6570\u7ea6\u675f\u8f93\u51fa\u7a7a\u95f4\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u8f83\u65e9\u63d0\u51fa\u81ea\u7136\u8bed\u8a00\u786c\u4ef6\u8bbe\u8ba1\u9700\u8981 pre-processing &#043; post-processing &#043; search\u3002 \u4ece\u201c\u8ba9 LLM \u5199\u4ee3\u7801\u201d\u8fc7\u6e21\u5230\u201c\u56f4\u7ed5 LLM \u5efa\u8bbe\u8ba1\u7a7a\u95f4\u63a2\u7d22\u5668\u201d\u3002 \u4e0e NVIDIA \u540e\u7eed agent\/verifier \u8def\u7ebf\u5728\u601d\u60f3\u4e0a\u76f8\u901a\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u505a\u5c0f\u578b\u7ec4\u5408\/\u65f6\u5e8f\u903b\u8f91\u8bbe\u8ba1\u63a2\u7d22&#xff0c;\u4e5f\u9002\u5408\u542f\u53d1\u4f01\u4e1a\u5185\u90e8 RTL \u751f\u6210\u5de5\u5177\u628a prompt \u6a21\u677f\u3001\u63a5\u53e3 schema\u3001lint\/\u4eff\u771f\u548c\u5019\u9009\u6392\u5e8f\u7ec4\u5408\u8d77\u6765\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u6846\u67b6\u4ecd\u504f demo \u548c\u5c0f\u8bbe\u8ba1&#xff0c;\u76ee\u6807\u51fd\u6570\u4e0e\u771f\u5b9e SoC \u7ea6\u675f\u76f8\u5dee\u8f83\u8fdc\u3002\u6ca1\u6709\u5f3a verifier\u3001testbench \u548c\u591a\u6587\u4ef6\u4e0a\u4e0b\u6587\u65f6&#xff0c;\u751f\u6210\u8d28\u91cf\u4e0a\u9650\u660e\u663e\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5b66\u672f\u8bba\u6587 \/ LLM4RTL benchmark<\/p>\n<h4>RTLLM&#xff1a;\u81ea\u7136\u8bed\u8a00\u5230 RTL \u7684\u5f00\u653e\u8bc4\u6d4b\u57fa\u51c6<\/h4>\n<p>arXiv 2308.05345\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04rm3du2dcv4g.png\" alt=\"RTLLM\uff1a\u81ea\u7136\u8bed\u8a00\u5230 RTL \u7684\u5f00\u653e\u8bc4\u6d4b\u57fa\u51c6 - RTLLM workflow \u622a\u56fe\" \/><\/p>\n<p>RTLLM workflow \u622a\u56fe \u00b7 PDF p.2 \u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a\u4ece prompt \u5230 RTL \u518d\u5230\u8bed\u6cd5\u3001\u529f\u80fd\u3001\u8d28\u91cf\u8bc4\u4f30\u7684\u6d41\u7a0b\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-0441nmzbeh3om.png\" alt=\"RTLLM\uff1a\u81ea\u7136\u8bed\u8a00\u5230 RTL \u7684\u5f00\u653e\u8bc4\u6d4b\u57fa\u51c6 - benchmark \u8868\u622a\u56fe\" \/><\/p>\n<p>benchmark \u8868\u622a\u56fe \u00b7 PDF p.3 \u8bba\u6587\u5185\u90e8\u8868\u683c\u5217\u51fa benchmark design \u63cf\u8ff0\u3001\u4ee3\u7801\u884c\u6570\u548c\u7535\u8def\u89c4\u6a21\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u65e9\u671f LLM4RTL \u5de5\u4f5c\u5404\u81ea\u4f7f\u7528\u5c0f\u6837\u4f8b&#xff0c;\u96be\u4ee5\u516c\u5e73\u6bd4\u8f83&#xff1b;\u5f88\u591a\u53ea\u770b\u529f\u80fd\u6b63\u786e&#xff0c;\u4e0d\u770b\u8bed\u6cd5\u3001\u8d28\u91cf\u548c\u89c4\u6a21\u3002RTLLM \u89e3\u51b3\u7684\u662f\u5982\u4f55\u4e3a\u81ea\u7136\u8bed\u8a00 RTL \u751f\u6210\u5efa\u7acb\u53ef\u91cf\u5316\u3001\u53ef\u590d\u7528 benchmark\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u5b9a\u4e49 syntax goal\u3001functionality goal\u3001design quality goal \u4e09\u5c42\u6e10\u8fdb\u76ee\u6807\u3002<\/li>\n<li>\u7ed9\u5b9a\u81ea\u7136\u8bed\u8a00\u4efb\u52a1\u540e\u81ea\u52a8\u751f\u6210 RTL&#xff0c;\u5e76\u901a\u8fc7\u4eff\u771f\/\u8d28\u91cf\u6307\u6807\u8fdb\u884c\u91cf\u5316\u8bc4\u4f30\u3002<\/li>\n<li>\u63d0\u51fa self-planning prompt \u6280\u672f&#xff0c;\u8ba9\u6a21\u578b\u5148\u89c4\u5212\u518d\u751f\u6210&#xff0c;\u63d0\u9ad8 GPT-3.5 \u5728 benchmark \u4e0a\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>arXiv \u6458\u8981\u660e\u786e\u4e09\u7c7b\u76ee\u6807&#xff1a;syntax\u3001functionality\u3001design quality\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8 workflow \u56fe\u5c55\u793a LLM \u8f93\u5165\u3001RTL \u751f\u6210\u3001\u81ea\u52a8\u8bc4\u6d4b\u548c PPA\/\u8d28\u91cf\u62a5\u544a\u7684\u95ed\u73af\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u8868\u683c\u5217\u51fa benchmark design \u63cf\u8ff0\u3001\u4ee3\u7801\u884c\u6570\u548c\u7535\u8def\u89c4\u6a21\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>RTLLM \u7684\u7406\u8bba\u8d21\u732e\u662f\u628a LLM4RTL \u4ece anecdotal demo \u62c9\u56de benchmark discipline\u3002\u5bf9\u786c\u4ef6\u8bbe\u8ba1\u6765\u8bf4&#xff0c;\u8bed\u6cd5\u6b63\u786e\u53ea\u662f\u6700\u4f4e\u95e8\u69db&#xff0c;\u529f\u80fd\u548c\u8d28\u91cf\u5fc5\u987b\u540c\u65f6\u8bc4\u4ef7\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a\u81ea\u7136\u8bed\u8a00 RTL \u751f\u6210\u8bc4\u6d4b\u4ece pass\/fail \u6269\u5c55\u5230\u591a\u76ee\u6807\u8d28\u91cf\u3002 self-planning \u8bf4\u660e prompt \u7ed3\u6784\u5bf9\u786c\u4ef6\u751f\u6210\u7ed3\u679c\u5f71\u54cd\u663e\u8457\u3002 \u4e3a\u540e\u7eed VerilogEval\u3001CVDP \u7b49\u66f4\u590d\u6742\u57fa\u51c6\u63d0\u4f9b\u6a2a\u5411\u53c2\u7167\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u7528\u4f5c\u5185\u90e8 RTL copilot \u7684\u5165\u95e8\u56de\u5f52\u96c6&#xff1a;\u6bcf\u6b21\u6a21\u578b\u3001prompt\u3001RAG \u6216 verifier \u6539\u52a8\u540e&#xff0c;\u5e94\u540c\u65f6\u8bb0\u5f55\u8bed\u6cd5\u3001\u529f\u80fd\u548c QoR \u53d8\u5316\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>benchmark \u8bbe\u8ba1\u4ecd\u6bd4\u771f\u5b9e IP \u5c0f\u5f97\u591a&#xff0c;\u4e14\u65e0\u6cd5\u8986\u76d6\u591a\u6587\u4ef6\u4f9d\u8d56\u3001\u590d\u6742 testbench\u3001\u7ea6\u675f\u3001CDC\/RDC\u3001\u4f4e\u529f\u8017\u548c\u7269\u7406\u53cd\u9988\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u7efc\u8ff0\u8bba\u6587 \/ ML for EDA \u5386\u53f2\u6846\u67b6<\/p>\n<h4>Machine Learning for EDA Survey&#xff1a;LLM \u4e4b\u524d\u7684 AI4EDA \u57fa\u7ebf<\/h4>\n<p>arXiv 2102.03357&#xff0c;TODAES survey\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-0424vxt0s4yn3.png\" alt=\"Machine Learning for EDA Survey\uff1aLLM \u4e4b\u524d\u7684 AI4EDA \u57fa\u7ebf - \u73b0\u4ee3\u82af\u7247\u8bbe\u8ba1\u6d41\u7a0b\u56fe\u622a\u56fe\" \/><\/p>\n<p>\u73b0\u4ee3\u82af\u7247\u8bbe\u8ba1\u6d41\u7a0b\u56fe\u622a\u56fe \u00b7 PDF p.7 \u7efc\u8ff0\u5185\u90e8\u56fe\u628a\u67b6\u6784\u3001\u903b\u8f91\u3001\u7269\u7406\u3001\u5236\u9020\u548c\u6d4b\u8bd5\u653e\u5165\u540c\u4e00\u8bbe\u8ba1\u6d41\u7a0b\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04cwfa0vgolkh.png\" alt=\"Machine Learning for EDA Survey\uff1aLLM \u4e4b\u524d\u7684 AI4EDA \u57fa\u7ebf - ML for HLS \u603b\u7ed3\u8868\u622a\u56fe\" \/><\/p>\n<p>ML for HLS \u603b\u7ed3\u8868\u622a\u56fe \u00b7 PDF p.10 \u7efc\u8ff0\u5185\u90e8\u8868\u683c\u5c55\u793a\u5178\u578b ML \u4efb\u52a1\u3001\u7b97\u6cd5\u548c\u5f15\u7528&#xff0c;\u662f\u5386\u53f2\u57fa\u7ebf\u8bc1\u636e\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u5728 LLM \u70ed\u6f6e\u4e4b\u524d&#xff0c;AI4EDA \u5df2\u7ecf\u8986\u76d6 HLS\u3001\u903b\u8f91\u7efc\u5408\u3001\u7269\u7406\u8bbe\u8ba1\u3001\u5149\u523b\u3001\u6a21\u62df\u3001\u9a8c\u8bc1\u6d4b\u8bd5\u7b49\u5f88\u591a\u4efb\u52a1\u3002\u8be5 survey \u7684\u4ef7\u503c\u662f\u7ed9\u51fa\u524d LLM \u65f6\u4ee3\u7684\u5168\u5c40\u5730\u56fe&#xff0c;\u907f\u514d\u628a AI4EDA \u8bef\u89e3\u4e3a 2023 \u4e4b\u540e\u624d\u51fa\u73b0\u7684 RTL \u751f\u6210\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6309 EDA hierarchy \u7cfb\u7edf\u7ec4\u7ec7 ML \u5e94\u7528&#xff1a;\u9ad8\u5c42\u7efc\u5408\u3001\u903b\u8f91\u7efc\u5408\u3001\u7269\u7406\u8bbe\u8ba1\u3001\u5236\u9020\u3001\u6a21\u62df\u3001\u9a8c\u8bc1\u6d4b\u8bd5\u7b49\u3002<\/li>\n<li>\u603b\u7ed3\u4e0d\u540c\u4efb\u52a1\u7684\u7279\u5f81\u3001\u6a21\u578b\u7c7b\u578b\u548c\u9002\u7528\u76ee\u6807&#xff0c;\u5982 surrogate prediction\u3001design space exploration\u3001optimization\u3002<\/li>\n<li>\u628a\u4f20\u7edf\u673a\u5668\u5b66\u4e60\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u56fe\u5b66\u4e60\u7b49\u65b9\u6cd5\u548c EDA \u9636\u6bb5\u5173\u8054\u8d77\u6765\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>arXiv \u6458\u8981\u8bf4\u660e\u8be5\u7efc\u8ff0\u6309 EDA hierarchy \u7ec4\u7ec7 ML for EDA studies\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u73b0\u4ee3\u82af\u7247\u8bbe\u8ba1\u6d41\u7a0b\u56fe\u5c55\u793a\u4ece\u67b6\u6784\u5230\u5236\u9020\u6d4b\u8bd5\u7684\u5b8c\u6574\u5c42\u7ea7\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u8868\u683c\u603b\u7ed3 HLS\u3001\u903b\u8f91\u7efc\u5408\u3001\u7269\u7406\u8bbe\u8ba1\u3001\u5149\u523b\u7b49\u4efb\u52a1\u4e2d\u7684 ML \u65b9\u6cd5\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>\u8be5 survey \u63d0\u4f9b\u7684\u7ed3\u8bba\u662f&#xff1a;LLM4EDA \u662f AI4EDA \u7684\u4e00\u4e2a\u65b0\u5206\u652f&#xff0c;\u800c\u4e0d\u662f\u5168\u90e8\u3002\u73b0\u4ee3 agent \u5fc5\u987b\u7ee7\u627f\u8fc7\u53bb\u4e8c\u5341\u591a\u5e74 ML for EDA \u7684\u9884\u6d4b\u3001\u4f18\u5316\u3001\u641c\u7d22\u548c\u7269\u7406\u7ea6\u675f\u7ecf\u9a8c\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u4f5c\u4e3a\u672c\u6587\u5916\u90e8\u8109\u7edc\u7684\u5386\u53f2\u57fa\u7ebf\u3002 \u8bf4\u660e\u540e\u7aef AI\u3001\u5236\u9020 AI \u548c\u9a8c\u8bc1 AI \u65e9\u5df2\u5b58\u5728\u3002 \u5e2e\u52a9\u533a\u5206\u201c\u8bed\u8a00\u6a21\u578b\u80fd\u529b\u201d\u548c\u201cEDA \u4f18\u5316\u80fd\u529b\u201d\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u505a AI4EDA \u6218\u7565\u65f6&#xff0c;\u5e94\u628a LLM agent \u4e0e\u4f20\u7edf ML predictor\u3001RL optimizer\u3001GPU kernel\u3001EDA heuristic \u7ec4\u5408&#xff0c;\u800c\u4e0d\u662f\u628a\u6240\u6709\u4efb\u52a1\u90fd\u4ea4\u7ed9\u901a\u7528\u5927\u6a21\u578b\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u7efc\u8ff0\u4e3b\u8981\u8986\u76d6 2021 \u5e74\u524d\u540e\u7684 ML \u5de5\u4f5c&#xff0c;\u5bf9 2023 \u4e4b\u540e\u7684 agent\u3001RAG\u3001long-context\u3001multimodal LLM \u8986\u76d6\u4e0d\u8db3\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u7efc\u8ff0\u8bba\u6587 \/ LLM4EDA \u5168\u5c40\u5730\u56fe<\/p>\n<h4>LLM for EDA Survey&#xff1a;\u4ece RTL \u751f\u6210\u5230\u5168\u6d41\u7a0b\u591a\u6a21\u6001 agent<\/h4>\n<p>arXiv 2508.20030\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04kmynnok15jf.png\" alt=\"LLM for EDA Survey\uff1a\u4ece RTL \u751f\u6210\u5230\u5168\u6d41\u7a0b\u591a\u6a21\u6001 agent - LLM-aided EDA agent \u56fe\u622a\u56fe\" \/><\/p>\n<p>LLM-aided EDA agent \u56fe\u622a\u56fe \u00b7 PDF p.2 \u7efc\u8ff0\u5185\u90e8\u56fe\u5c55\u793a LLM agent \u5982\u4f55\u6a2a\u8de8 full-flow\u3001\u591a\u6a21\u6001\u548c\u81ea\u52a8\u5316\u4efb\u52a1\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04suvwi2dqkdh.png\" alt=\"LLM for EDA Survey\uff1a\u4ece RTL \u751f\u6210\u5230\u5168\u6d41\u7a0b\u591a\u6a21\u6001 agent - \u4efb\u52a1\u6846\u67b6\u56fe\/\u8868\u622a\u56fe\" \/><\/p>\n<p>\u4efb\u52a1\u6846\u67b6\u56fe\/\u8868\u622a\u56fe \u00b7 PDF p.2 \u7efc\u8ff0\u5185\u90e8\u56fe\u8868\u5c06 HLS \u4fee\u590d\u3001RAG\u3001PPA \u4f18\u5316\u548c\u9a8c\u8bc1\u7eb3\u5165\u540c\u4e00 taxonomy\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>LLM4EDA \u7684\u8bba\u6587\u8fc5\u901f\u5206\u6563\u5230 RTL \u751f\u6210\u3001HLS \u4fee\u590d\u3001\u9a8c\u8bc1\u3001\u811a\u672c\u3001\u7269\u7406\u8bbe\u8ba1\u3001\u5236\u9020\u548c\u591a\u6a21\u6001 agent\u3002\u8be5 survey \u7684\u95ee\u9898\u662f\u5982\u4f55\u628a\u8fd9\u4e9b\u65b9\u5411\u653e\u56de\u4e00\u4e2a\u7edf\u4e00 taxonomy&#xff0c;\u5224\u65ad\u54ea\u4e9b\u53ea\u662f\u6587\u672c\u751f\u6210&#xff0c;\u54ea\u4e9b\u63a5\u8fd1\u5de5\u5177\u95ed\u73af\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u6309\u82af\u7247\u8bbe\u8ba1\u9636\u6bb5\u548c LLM \u5e94\u7528\u7c7b\u578b\u7ec4\u7ec7\u5de5\u4f5c&#xff1a;design\u3001testing\u3001optimization\u3001agentic workflow\u3002<\/li>\n<li>\u8ba8\u8bba LLM \u4e0e HLS\u3001RTL\u3001EDA \u811a\u672c\u3001\u9a8c\u8bc1\u3001\u7269\u7406\u8bbe\u8ba1\u3001\u591a\u6a21\u6001\u4fe1\u606f\u4e4b\u95f4\u7684\u63a5\u53e3\u3002<\/li>\n<li>\u5f3a\u8c03\u5b8c\u6574 specification-to-silicon \u9700\u8981\u8de8\u81ea\u7136\u8bed\u8a00\u3001\u9ad8\u5c42\u8bed\u8a00\u3001RTL \u548c\u7269\u7406\u5b9e\u73b0\u4fdd\u6301\u8bed\u4e49\u4e00\u81f4\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u8bba\u6587\u5185\u90e8\u56fe\u5c55\u793a LLM-aided EDA agent \u8de8 full flow \u7684\u591a\u6a21\u6001\u80fd\u529b\u3002<\/li>\n<li>\u8bba\u6587\u5185\u90e8\u8868\u683c\/\u6846\u67b6\u56fe\u5c55\u793a HLS \u4fee\u590d\u3001RAG\u3001PPA optimization\u3001equivalence verification \u7b49\u4efb\u52a1\u7c7b\u578b\u3002<\/li>\n<li>\u4e0e NVIDIA \u8bba\u6587\u8c31\u7cfb\u76f8\u4e92\u5370\u8bc1&#xff1a;2025-2026 \u7684\u91cd\u70b9\u5df2\u7ecf\u4ece\u5355\u6b21\u751f\u6210\u8f6c\u5411 agent\u3001verifier \u548c\u957f\u4e0a\u4e0b\u6587\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>\u8be5 survey \u7684\u6838\u5fc3\u7ed3\u8bba\u662f&#xff0c;LLM4EDA \u7684\u7ec8\u5c40\u4e0d\u662f\u5355\u70b9\u751f\u6210\u5668&#xff0c;\u800c\u662f\u8de8\u8868\u793a\u3001\u8de8\u5de5\u5177\u3001\u8de8\u9636\u6bb5\u7684\u534f\u540c agent\u3002\u771f\u6b63\u96be\u70b9\u662f\u8bed\u4e49\u4fdd\u6301\u548c\u53ef\u9a8c\u8bc1\u6027\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u628a\u5206\u6563\u7684 LLM4EDA \u5de5\u4f5c\u5f52\u5165 full-flow \u56fe\u8c31\u3002 \u5f3a\u8c03\u591a\u6a21\u6001\u548c\u5de5\u5177\u4ea4\u4e92&#xff0c;\u800c\u4e0d\u662f\u53ea\u5f3a\u8c03 chat\u3002 \u4e3a\u672c\u6587\u7684\u4e94\u5c42\u80fd\u529b\u6808\u63d0\u4f9b\u5916\u90e8\u7efc\u8ff0\u652f\u6491\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u4f5c\u4e3a\u6280\u672f\u8def\u7ebf\u5bf9\u7167\u8868&#xff1a;\u4f01\u4e1a\u53ef\u4ee5\u7528\u5b83\u68c0\u67e5\u81ea\u5df1\u662f\u5426\u53ea\u505a\u4e86\u95ee\u7b54\/RAG&#xff0c;\u8fd8\u662f\u5df2\u7ecf\u8986\u76d6\u811a\u672c\u6267\u884c\u3001\u9519\u8bef\u4fee\u590d\u3001\u9a8c\u8bc1\u53cd\u9988\u3001PPA \u53cd\u9988\u548c\u7269\u7406\u8bc1\u636e\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u7efc\u8ff0\u6027\u8d28\u51b3\u5b9a\u5176\u66f4\u591a\u662f\u7ed3\u6784\u5316\u5f52\u7eb3&#xff0c;\u4e0d\u662f\u5355\u4e00\u53ef\u590d\u73b0\u5b9e\u9a8c\u7cfb\u7edf\u3002\u5177\u4f53\u6027\u80fd\u4ecd\u9700\u56de\u5230\u5404\u539f\u8bba\u6587\u548c\u5de5\u5177\u94fe\u9a8c\u8bc1\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u7814\u8ba8\u4f1a\u62a5\u544a \/ AI4EDA \u7814\u7a76\u8bae\u7a0b<\/p>\n<h4>NSF AI4EDA Workshop Report&#xff1a;\u4ece\u8bba\u6587\u7ade\u4e89\u8d70\u5411\u751f\u6001\u5efa\u8bbe<\/h4>\n<p>arXiv 2601.14541&#xff0c;IEEE Circuits and Systems Magazine 2026 accepted version\u3002 \u00b7 source<\/p>\n<p><img decoding=\"async\" src=\"2026-08-04g5p0mah505n.png\" alt=\"NSF AI4EDA Workshop Report\uff1a\u4ece\u8bba\u6587\u7ade\u4e89\u8d70\u5411\u751f\u6001\u5efa\u8bbe - \u62a5\u544a\u6b63\u6587\u5c40\u90e8\u622a\u56fe\" \/><\/p>\n<p>\u62a5\u544a\u6b63\u6587\u5c40\u90e8\u622a\u56fe \u00b7 PDF p.2 \u62a5\u544a\u65e0\u7a33\u5b9a Figure\/Table caption&#xff0c;\u672c\u56fe\u622a\u53d6\u6b63\u6587\u4e2d\u5173\u4e8e LLM agent\u3001\u9a8c\u8bc1\u548c workforce \u7684\u5efa\u8bae\u6bb5\u843d\u3002<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u5f53 AI4EDA \u8fdb\u5165 full-flow\u3001foundation model \u548c\u4ea7\u4e1a\u843d\u5730\u9636\u6bb5&#xff0c;\u5355\u4e2a benchmark \u6216\u5355\u7bc7\u8bba\u6587\u5df2\u65e0\u6cd5\u89e3\u51b3\u6570\u636e\u3001\u7b97\u529b\u3001\u5f00\u653e\u5de5\u5177\u3001\u4eba\u624d\u3001\u534f\u4f5c\u673a\u5236\u7684\u95ee\u9898\u3002NSF \u62a5\u544a\u5173\u6ce8\u7684\u662f\u56fd\u5bb6\u7ea7\/\u751f\u6001\u7ea7\u7814\u7a76\u8bae\u7a0b\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u628a workshop \u8ba8\u8bba\u5f52\u7eb3\u4e3a\u56db\u4e2a\u4e3b\u9898&#xff1a;physical synthesis\/DFM\u3001HLS\/LLS\u3001AI toolbox for optimization\/design\u3001test\/verification\u3002<\/li>\n<li>\u63d0\u51fa\u4fc3\u8fdb AI\/EDA collaboration\u3001\u6295\u8d44 EDA foundation AI\u3001\u5efa\u7acb robust data infrastructure\u3001\u6269\u5c55 scalable compute\u3001\u57f9\u517b workforce\u3002<\/li>\n<li>\u628a AI4EDA \u4ece\u7b97\u6cd5\u95ee\u9898\u4e0a\u5347\u4e3a\u57fa\u7840\u8bbe\u65bd\u3001\u6559\u80b2\u548c\u5f00\u653e\u751f\u6001\u95ee\u9898\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>arXiv \u6458\u8981\u660e\u786e workshop \u4e8e 2024 \u5e74 12 \u6708 NeurIPS \u671f\u95f4\u4e3e\u884c&#xff0c;\u5e76\u5217\u51fa\u56db\u5927\u4e3b\u9898\u3002<\/li>\n<li>\u6458\u8981\u7ed9\u51fa recommendations&#xff1a;AI\/EDA collaboration\u3001foundational AI for EDA\u3001robust data infrastructure\u3001scalable compute\u3001workforce development\u3002<\/li>\n<li>\u672c\u5730\u622a\u56fe\u622a\u53d6\u4e86\u62a5\u544a\u6b63\u6587\u4e2d\u5bf9 LLM agent\u3001open EDA\u3001\u9a8c\u8bc1\u548c workforce \u7684\u5efa\u8bae\u6bb5\u843d\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>NSF \u62a5\u544a\u7684\u7ed3\u8bba\u662f AI4EDA \u6b63\u5728\u4ece\u201c\u6a21\u578b\u80fd\u4e0d\u80fd\u505a\u67d0\u4e2a\u4efb\u52a1\u201d\u8f6c\u5411\u201c\u884c\u4e1a\u80fd\u4e0d\u80fd\u5efa\u8bbe\u5171\u4eab\u6570\u636e\u3001\u5f00\u653e\u5de5\u5177\u3001\u53ef\u4fe1\u8bc4\u6d4b\u548c\u4eba\u624d\u4f53\u7cfb\u201d\u3002\u8fd9\u662f\u4ea7\u4e1a\u5316\u62d0\u70b9\u7684\u91cd\u8981\u6807\u5fd7\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u8986\u76d6\u5b66\u672f\u3001\u4ea7\u4e1a\u548c\u6559\u80b2\u4e09\u7c7b\u95ee\u9898\u3002 \u628a verifier\u3001data infrastructure\u3001compute infrastructure \u89c6\u4e3a\u540c\u7b49\u91cd\u8981\u7684\u957f\u671f\u8d44\u4ea7\u3002 \u4e0e NVIDIA Trace2Skill\/ACE-RTL \u7684\u65b9\u5411\u4e00\u81f4&#xff1a;\u5de5\u5177\u53cd\u9988\u548c\u6280\u80fd\u79ef\u7d2f\u662f\u6838\u5fc3\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u4f01\u4e1a\u5185\u90e8\u843d\u5730 AI4EDA \u65f6&#xff0c;\u5e94\u628a roadmap \u62c6\u6210\u6570\u636e\u6cbb\u7406\u3001\u5de5\u5177 API\u3001\u9a8c\u8bc1 oracle\u3001\u7b97\u529b\u8c03\u5ea6\u3001\u6a21\u578b\/agent\u3001\u8bc4\u6d4b\u548c\u4eba\u624d\u8bad\u7ec3&#xff0c;\u800c\u4e0d\u662f\u53ea\u91c7\u8d2d\u6216\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u62a5\u544a\u662f\u7814\u7a76\u8bae\u7a0b&#xff0c;\u4e0d\u662f benchmark \u6216\u5546\u4e1a\u4ea7\u54c1\u3002\u5176\u5224\u65ad\u9700\u8981\u7ed3\u5408\u5177\u4f53\u5de5\u5177\u94fe\u548c\u7ec4\u7ec7\u80fd\u529b\u843d\u5730\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4ea7\u4e1a\u65b9\u6848 \/ \u5546\u4e1a AI EDA suite<\/p>\n<h4>Synopsys.ai&#xff1a;\u5546\u4e1a EDA AI \u5e73\u53f0\u5316\u8def\u7ebf<\/h4>\n<p>Synopsys \u5b98\u65b9 Synopsys.ai \u9875\u9762\u3002 \u00b7 source<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u4ea7\u4e1a\u754c\u7684\u95ee\u9898\u4e0d\u662f\u67d0\u4e2a benchmark \u7684 pass&#064;k&#xff0c;\u800c\u662f\u5982\u4f55\u5728\u6570\u5b57\u8bbe\u8ba1\u3001\u6a21\u62df\u3001\u9a8c\u8bc1\u3001DFT\u3001\u5de5\u827a\u826f\u7387\u548c\u7cfb\u7edf\u67b6\u6784\u4e2d\u6301\u7eed\u964d\u4f4e TAT\u3001\u63d0\u5347 PPA \u548c\u51cf\u5c11\u4eba\u5de5\u8fed\u4ee3\u3002Synopsys.ai \u4ee3\u8868\u5546\u4e1a EDA \u4ece\u5355\u5de5\u5177\u667a\u80fd\u5316\u8d70\u5411\u5e73\u53f0\u5316\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u56f4\u7ed5 DSO.ai\u3001VSO.ai\u3001TSO.ai\u3001ASO.ai\u30013DSO.ai \u7b49\u4e0d\u540c\u529f\u80fd&#xff0c;\u628a AI \u5d4c\u5165 design\u3001verification\u3001test\u3001analog\u30013DIC \u548c manufacturing\u3002<\/li>\n<li>\u5b98\u65b9\u9875\u9762\u540c\u65f6\u5f3a\u8c03 AI-powered EDA\u3001Generative AI \u548c Agentic AI&#xff0c;\u8bf4\u660e\u4ea7\u54c1\u53d9\u4e8b\u4ece\u4f18\u5316\u5668\u6269\u5c55\u5230 copilot \u548c multi-agent workflow\u3002<\/li>\n<li>\u7528\u5386\u53f2\u9879\u76ee\u6570\u636e\u3001EDA \u5de5\u5177\u5185\u751f telemetry \u548c\u6d41\u7a0b\u77e5\u8bc6\u63d0\u4f9b floorplan\u3001timing\u3001power\u3001coverage\u3001yield \u7b49\u5efa\u8bae\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u5b98\u65b9\u9875\u9762\u58f0\u660e silicon lifecycle \u6700\u9ad8 30% productivity gains&#xff0c;\u4ee5\u53ca next-generation chips development cycles 5X \u52a0\u901f\u3002<\/li>\n<li>\u9875\u9762\u5217\u51fa digital design\u3001analog\u3001verification\u3001system architects\u3001DFT\u3001process\/yield \u7b49\u529f\u80fd\u5165\u53e3\u3002<\/li>\n<li>\u9875\u9762\u628a generative AI \u548c agentic AI \u5355\u5217\u4e3a\u4ea7\u54c1\u80fd\u529b&#xff0c;\u53cd\u6620\u5546\u4e1a EDA \u5df2\u8fdb\u5165\u5e73\u53f0\u5316\u7ade\u4e89\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>Synopsys.ai \u7684\u7406\u8bba\u610f\u4e49\u5728\u4e8e&#xff1a;\u4ea7\u4e1a AI4EDA \u6700\u7ec8\u4f1a\u5d4c\u5165\u65e2\u6709 signoff \u5de5\u5177\u548c\u6570\u636e\u95ed\u73af&#xff0c;\u800c\u4e0d\u662f\u4ee5\u72ec\u7acb chatbot \u5b58\u5728\u3002AI \u7684\u4ef7\u503c\u7531\u5de5\u5177\u94fe\u4e2d\u7684\u53ef\u6267\u884c\u51b3\u7b56\u548c\u751f\u4ea7\u7387\u6307\u6807\u8861\u91cf\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u8986\u76d6 silicon lifecycle \u591a\u9636\u6bb5&#xff0c;\u4e0d\u5c40\u9650 RTL\u3002 \u628a AI \u4f18\u5316\u3001\u6570\u636e\u5206\u6790\u3001\u751f\u6210\u5f0f\u52a9\u624b\u548c agent workflow \u653e\u5230\u540c\u4e00\u5546\u4e1a\u53d9\u4e8b\u3002 \u5f3a\u8c03\u751f\u4ea7\u7387\u548c\u5f00\u53d1\u5468\u671f&#xff0c;\u800c\u4e0d\u662f\u5b66\u672f benchmark\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u5bf9\u4f01\u4e1a\u6700\u73b0\u5b9e\u7684\u542f\u793a\u662f&#xff1a;AI4EDA \u9700\u8981\u63a5\u5165\u73b0\u6709 EDA \u5de5\u5177\u6570\u636e\u3001\u9879\u76ee\u5386\u53f2\u3001\u7b7e\u6838\u6307\u6807\u548c\u6743\u9650\u4f53\u7cfb&#xff1b;\u5355\u72ec\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\u65e0\u6cd5\u66ff\u4ee3\u5546\u4e1a flow \u91cc\u7684\u95ed\u73af\u8d44\u4ea7\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u5546\u4e1a\u9875\u9762\u7684\u6307\u6807\u9700\u8981\u7ed3\u5408\u5ba2\u6237\u3001\u8282\u70b9\u3001\u8bbe\u8ba1\u89c4\u6a21\u3001baseline \u548c\u5de5\u5177\u7248\u672c\u89e3\u91ca\u3002\u516c\u5f00\u6750\u6599\u901a\u5e38\u4e0d\u63d0\u4f9b\u5b8c\u6574\u53ef\u590d\u73b0\u5b9e\u9a8c\u7ec6\u8282\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4ea7\u4e1a\u65b9\u6848 \/ \u5546\u4e1a RTL-to-GDS AI \u4f18\u5316<\/p>\n<h4>Cadence Cerebrus \/ AI Studio&#xff1a;PPA \u4f18\u5316\u4e0e agentic SoC \u8bbe\u8ba1\u5e73\u53f0<\/h4>\n<p>Cadence \u5b98\u65b9 Cerebrus Intelligent Chip Explorer \u9875\u9762\u3002 \u00b7 source<\/p>\n<h5>\u95ee\u9898\u5b9a\u4e49<\/h5>\n<p>\u5148\u8fdb SoC \u7684 PPA closure \u9700\u8981\u5927\u91cf block \u7ea7\u5e76\u884c\u63a2\u7d22\u3001flow \u53c2\u6570\u8c03\u4f18\u548c\u7ed3\u679c\u5206\u6790\u3002Cerebrus \u89e3\u51b3\u7684\u662f\u5982\u4f55\u628a full-flow reinforcement learning\u3001LLM \u548c\u5206\u5e03\u5f0f\u8ba1\u7b97\u5d4c\u5165 RTL-to-GDS \u4f18\u5316\u3002<\/p>\n<h5>\u65b9\u6cd5 \/ \u67b6\u6784\u62c6\u89e3<\/h5>\n<ul>\n<li>\u8bbe\u8ba1\u56e2\u961f\u6307\u5b9a PPA \u76ee\u6807&#xff0c;Cerebrus \u81ea\u52a8\u4f18\u5316\u6570\u5b57\u8bbe\u8ba1 flow&#xff0c;\u4ece RTL \u5230 GDS \u641c\u7d22\u66f4\u4f18\u5b9e\u73b0\u65b9\u6848\u3002<\/li>\n<li>\u4f7f\u7528 full-flow reinforcement learning&#xff0c;\u5e76\u5728 Cadence JedAI\/Cerebrus AI Studio \u4e2d\u5f15\u5165 LLM \u80fd\u529b\u3002<\/li>\n<li>\u652f\u6301 multi-block\u3001multi-user\u3001\u5206\u5e03\u5f0f\u8ba1\u7b97\u548c designer cockpit&#xff0c;\u8ba9\u5de5\u7a0b\u5e08\u5206\u6790\u7ed3\u679c\u5e76\u4fdd\u6301\u63a7\u5236\u3002<\/li>\n<\/ul>\n<h5>\u8bc1\u636e\u94fe<\/h5>\n<ul>\n<li>\u5b98\u65b9\u9875\u9762\u79f0 Cerebrus \u901a\u8fc7 AI-driven automated approach \u6539\u5584 PPA \u548c engineering productivity\u3002<\/li>\n<li>\u9875\u9762\u660e\u786e\u63d0\u5230 full-flow reinforcement learning technology \u548c LLM capabilities\u3002<\/li>\n<li>\u9875\u9762\u79f0 Cerebrus AI Studio \u652f\u6301 multi-block\u3001multi-user design&#xff0c;\u5e76\u52a0\u901f SoC delivery 5X to 10X&#xff1b;\u5ba2\u6237\u6bb5\u843d\u5217\u51fa MediaTek die area\/power \u548c Renesas TNS \u7b49\u6539\u5584\u6848\u4f8b\u3002<\/li>\n<\/ul>\n<h5>\u673a\u5236\u6846\u67b6\u56fe<\/h5>\n<h5>\u7406\u8bba\u7ed3\u8bba<\/h5>\n<p>Cadence Cerebrus \u7684\u7ed3\u8bba\u662f\u5546\u4e1a AI4EDA \u7684\u4e2d\u5fc3\u6307\u6807\u662f PPA\/TAT&#xff0c;\u800c\u4e0d\u662f\u6a21\u578b\u80fd\u529b\u5c55\u793a\u3002AI \u5728\u8fd9\u91cc\u662f flow optimizer \u548c\u8bbe\u8ba1\u9a7e\u9a76\u8231&#xff0c;\u4e0d\u662f\u72ec\u7acb\u5bf9\u8bdd\u6846\u3002<\/p>\n<h5>\u6210\u719f\u5ea6\u4e0e\u98ce\u9669\u753b\u50cf<\/h5>\n<table>\n<tr>\u5173\u952e\u4eae\u70b9\u76f4\u63a5\u7ed1\u5b9a RTL-to-GDS \u548c PPA closure\u3002 \u628a RL\u3001LLM\u3001\u5206\u5e03\u5f0f\u8ba1\u7b97\u548c\u4eba\u673a\u4ea4\u4e92 cockpit \u5408\u5e76\u3002 \u5ba2\u6237\u6848\u4f8b\u7ed9\u51fa\u66f4\u63a5\u8fd1\u4ea7\u4e1a\u91c7\u8d2d\u51b3\u7b56\u7684\u4fe1\u53f7\u3002<\/tr>\n<tbody>\n<tr>\n<td>\u5de5\u7a0b\u542f\u793a<\/td>\n<td>\u9002\u5408\u590d\u6742 block \u6216 SoC \u591a block \u5e76\u884c\u5b9e\u73b0\u63a2\u7d22\u3002\u4f01\u4e1a\u5185\u90e8\u81ea\u7814 agent \u65f6\u5e94\u5b66\u4e60\u5176\u95ed\u73af\u5f62\u6001&#xff1a;\u76ee\u6807\u5b9a\u4e49\u3001run \u7ba1\u7406\u3001\u7ed3\u679c\u5bf9\u6bd4\u3001\u6a21\u578b\u590d\u7528\u548c\u5de5\u7a0b\u5e08\u53ef\u63a7\u6027\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u7528\u8fb9\u754c<\/td>\n<td>\u516c\u5f00\u8d44\u6599\u5c5e\u4e8e\u5546\u4e1a\u9875\u9762&#xff0c;\u5177\u4f53\u7b97\u6cd5\u3001\u8bad\u7ec3\u6570\u636e\u3001baseline \u548c\u9002\u7528\u8bbe\u8ba1\u6761\u4ef6\u4e0d\u53ef\u5b8c\u5168\u590d\u73b0\u3002\u5ba2\u6237\u6848\u4f8b\u9700\u8981\u7ed3\u5408\u8282\u70b9\u3001\u76ee\u6807\u3001\u8bbe\u8ba1\u5927\u5c0f\u548c flow \u914d\u7f6e\u89e3\u8bfb\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u5168\u7403\u65b9\u6cd5\u8c31\u7cfb&#xff1a;\u516b\u7c7b\u8def\u7ebf&#xff0c;\u4e0d\u662f\u4e00\u4e2a\u8d5b\u9053<\/h3>\n<p>\u628a NVIDIA \u548c\u975e NVIDIA \u5de5\u4f5c\u653e\u5728\u540c\u4e00\u4e2a\u5750\u6807\u7cfb\u91cc&#xff0c;AI4EDA \u81f3\u5c11\u6709\u516b\u6761\u65b9\u6cd5\u7ebf\u3002\u5b83\u4eec\u7684\u8f93\u5165\u5bf9\u8c61\u3001\u53cd\u9988\u4fe1\u53f7\u3001\u6210\u719f\u5ea6\u548c\u843d\u5730\u8def\u5f84\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n<h4>RL\/GNN \u7269\u7406\u8bbe\u8ba1<\/h4>\n<p>AlphaChip \u628a\u5b8f\u5355\u5143\u5e03\u5c40\u8f6c\u6210\u5e8f\u5217\u51b3\u7b56&#xff0c;GNN \u7f16\u7801 netlist&#xff0c;RL \u5728\u4ee3\u7406 reward \u548c\u5de5\u4e1a\u5de5\u5177\u9a8c\u8bc1\u4e4b\u95f4\u641c\u7d22\u3002<\/p>\n<h4>GPU\/\u6570\u503c\u4f18\u5316<\/h4>\n<p>DREAMPlace \u7684\u542f\u793a\u662f&#xff1a;AI4EDA \u4e0d\u53ea\u7b49\u4e8e LLM\u3002\u628a placement \u5199\u6210\u5f20\u91cf\u4f18\u5316\u5e76\u642c\u4e0a GPU&#xff0c;\u672c\u8eab\u5c31\u662f\u4e00\u79cd AI \u65f6\u4ee3\u7684 EDA \u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n<h4>\u5f00\u653e flow \u4e0e\u5f00\u653e\u6570\u636e<\/h4>\n<p>OpenROAD\u3001EDA Corpus\u3001CircuitNet\u3001ForgeEDA \u89e3\u51b3\u7684\u662f\u53ef\u590d\u73b0\u95ee\u9898&#xff1a;\u6ca1\u6709\u5f00\u653e\u5de5\u5177\u3001\u5f00\u653e\u6570\u636e\u548c\u5f00\u653e\u8bc4\u6d4b&#xff0c;\u9886\u57df\u6a21\u578b\u5f88\u96be\u5f62\u6210\u516c\u5171\u8fdb\u6b65\u3002<\/p>\n<h4>\u81ea\u7136\u8bed\u8a00\u5230 RTL<\/h4>\n<p>Chip-Chat\u3001ChipGPT\u3001RTLLM\u3001OpenLLM-RTL \u628a\u201c\u5de5\u7a0b\u5e08\u89c4\u683c\u201d\u8f6c\u6210 HDL&#xff0c;\u4f46\u8fdb\u6b65\u7ebf\u7d22\u4ece prompt \u5f88\u5feb\u8f6c\u5411 testbench\u3001formal\u3001\u6570\u636e\u8d28\u91cf\u548c\u8bbe\u8ba1\u8d28\u91cf\u6307\u6807\u3002<\/p>\n<h4>\u9a8c\u8bc1\u4e0e agent<\/h4>\n<p>FVEval\u3001AssertionForge\u3001PRO-V-R1\u3001CVDP\u3001Trace2Skill \u5bf9\u5e94\u7684\u662f\u9a8c\u8bc1\u4fa7&#xff1a;assertion\u3001coverage\u3001fault detection\u3001hidden tests \u548c skill evolution\u3002<\/p>\n<h4>\u957f\u4e0a\u4e0b\u6587 benchmark<\/h4>\n<p>2026 \u5e74 RTL-BenchLS \u628a\u89c4\u6a21\u63a8\u5230 10,000&#043; formally verified Verilog designs&#xff0c;\u5e76\u5f15\u5165 round-trip\u3001masked-content\u3001repository-issue reasoning\u3002<\/p>\n<h4>\u5546\u4e1a\u5e73\u53f0\u5316<\/h4>\n<p>Synopsys\u3001Cadence\u3001Siemens \u7684\u8def\u7ebf\u4e0d\u662f\u8bba\u6587\u5f0f\u5355\u70b9 SOTA&#xff0c;\u800c\u662f\u628a AI \u653e\u8fdb\u65e2\u6709 signoff engine\u3001debug database\u3001tool API \u548c\u5ba2\u6237 flow\u3002<\/p>\n<h4>\u5236\u9020\u4e0e fab AI<\/h4>\n<p>cuLitho\u3001TSMC fab AI\u3001process simulation\u3001inspection \u548c FabTwin \u663e\u793a AI4EDA \u6b63\u5728\u5411 AI4Manufacturing \u5ef6\u4f38&#xff0c;\u65b9\u6cd5\u66f4\u504f HPC\u3001surrogate model\u3001vision \u548c digital twin\u3002<\/p>\n<h4>\u7efc\u8ff0\u4e0e\u516c\u5171\u8def\u7ebf\u56fe<\/h4>\n<p>ML\/LLM for EDA surveys \u4e0e NSF workshop \u628a\u5355\u70b9\u8bba\u6587\u5f52\u5165\u66f4\u5927\u7684\u5b66\u79d1\u5de5\u7a0b&#xff1a;\u6570\u636e\u57fa\u7840\u8bbe\u65bd\u3001\u53ef\u6269\u5c55 compute\u3001verification\u3001DFM \u548c\u4eba\u624d\u57f9\u517b\u3002<\/p>\n<h3>\u4ea7\u4e1a\u5e73\u53f0\u8865\u5145&#xff1a;\u5546\u4e1a EDA \u516c\u53f8\u7684\u5165\u53e3\u4e0d\u540c<\/h3>\n<p>\u524d\u9762\u7684\u5168\u7403\u7ae0\u8282\u5df2\u7ecf\u6df1\u8bfb Synopsys \u548c Cadence\u3002\u8fd9\u91cc\u8865\u4e0a Siemens\u3001\u5236\u9020\u4fa7 NVIDIA\/TSMC&#xff0c;\u5e76\u628a\u4e09\u5927 EDA \u5382\u5546\u7684\u8def\u7ebf\u653e\u5728\u540c\u4e00\u5f20\u8868\u91cc\u3002<\/p>\n<table>\n<tr>\u516c\u53f8\/\u673a\u6784AI \u5165\u53e3\u65b9\u6cd5\u91cd\u5fc3\u4e0e\u8bba\u6587\u7ebf\u7684\u5173\u7cfb<\/tr>\n<tbody>\n<tr>\n<td>Synopsys.ai<\/td>\n<td>Digital\u3001Analog\u3001Verification\u3001DFT\u3001Process\/Yield \u5168\u751f\u547d\u5468\u671f<\/td>\n<td>AI-powered EDA\u3001GenAI Copilot\u3001Agentic AI\u3001DSO\/ASO\/VSO\/TSO\/3DSO<\/td>\n<td>\u628a\u5355\u70b9\u7b97\u6cd5\u63a5\u5165\u5546\u4e1a signoff \u4e0e\u5ba2\u6237 flow&#xff0c;\u5f3a\u8c03\u751f\u4ea7\u529b\u3001TAT \u548c first-pass silicon\u3002<\/td>\n<\/tr>\n<tr>\n<td>Cadence<\/td>\n<td>Cerebrus\u3001AI Studio\u3001ChipStack\u3001AgentStack<\/td>\n<td>RL\/LLM PPA \u4f18\u5316\u3001\u9a8c\u8bc1 super agent\u3001multi-agent orchestration\u3001NVIDIA Nemotron\/OpenShell<\/td>\n<td>\u6700\u63a5\u8fd1 NVIDIA agent \u8bba\u6587\u7684\u4ea7\u4e1a\u5316\u53d9\u4e8b&#xff1a;\u4ece\u5de5\u5177\u52a9\u624b\u8d70\u5411\u201c\u865a\u62df\u5de5\u7a0b\u5e08\u201d\u3002<\/td>\n<\/tr>\n<tr>\n<td>Siemens EDA<\/td>\n<td>Solido\u3001Aprisa\u3001Fuse EDA AI System<\/td>\n<td>\u6a21\u62df\/\u6df7\u5408\u4fe1\u53f7\u3001\u5e93\u8868\u5f81\u3001\u4eff\u771f\u7ed3\u679c\u603b\u7ed3\u3001RTL-to-GDS agent\u3001\u81ea\u7136\u8bed\u8a00\u5de5\u5177\u4ea4\u4e92<\/td>\n<td>\u8bf4\u660e AI4EDA \u4e0d\u53ea\u5728\u6570\u5b57 RTL \u548c\u9a8c\u8bc1&#xff0c;custom IC \u4e0e characterization \u540c\u6837\u662f\u9ad8\u4ef7\u503c\u5165\u53e3\u3002<\/td>\n<\/tr>\n<tr>\n<td>NVIDIA \/ TSMC \/ ASML \/ IMEC<\/td>\n<td>cuLitho\u3001computational lithography\u3001process\/fab AI<\/td>\n<td>GPU \u52a0\u901f\u3001AI surrogate\u3001defect inspection\u3001advanced process control\u3001FabTwin<\/td>\n<td>\u628a AI4EDA \u63a8\u5230\u5236\u9020\u4fa7&#xff1a;\u4ece\u8bbe\u8ba1\u81ea\u52a8\u5316\u6269\u5c55\u5230 mask\u3001process window\u3001yield \u548c fab operations\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u6a2a\u7eb5\u4ea4\u6c47&#xff1a;\u4e3a\u4ec0\u4e48\u8fd9\u4e9b\u8def\u7ebf\u6700\u540e\u4f1a\u5408\u6d41<\/h3>\n<p>NVIDIA \u8bba\u6587\u7ebf\u3001Google \u7269\u7406\u8bbe\u8ba1\u7ebf\u3001OpenROAD \u5f00\u653e\u57fa\u7840\u8bbe\u65bd\u7ebf\u3001Synopsys\/Cadence\/Siemens \u5546\u4e1a\u5e73\u53f0\u7ebf&#xff0c;\u8868\u9762\u4e0a\u5404\u505a\u5404\u7684&#xff0c;\u5e95\u5c42\u5176\u5b9e\u5728\u89e3\u51b3\u540c\u4e00\u7ec4\u7ea6\u675f\u3002<\/p>\n<table>\n<tr>\u5171\u540c\u7ea6\u675fNVIDIA \u7ed9\u51fa\u7684\u7b54\u6848\u5168\u7403\u6a2a\u8f74\u7ed9\u51fa\u7684\u7b54\u6848\u6211\u7684\u5224\u65ad<\/tr>\n<tbody>\n<tr>\n<td>\u6b63\u786e\u6027\u4e0d\u80fd\u9760\u6a21\u578b\u81ea\u8bc4<\/td>\n<td>compiler\u3001simulator\u3001formal\u3001STA\u3001hidden verifier\u3001trace feedback<\/td>\n<td>RTLLM\/OpenLLM-RTL\/RTL-BenchLS\u3001OpenROAD flow\u3001industrial signoff engines<\/td>\n<td>\u672a\u6765 AI4EDA \u5e73\u53f0\u7684\u6838\u5fc3\u8d44\u4ea7\u4e0d\u662f prompt&#xff0c;\u800c\u662f\u53ef\u6267\u884c\u9a8c\u8bc1\u73af\u5883\u548c\u5931\u8d25\u8f68\u8ff9\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u8bbe\u8ba1\u5bf9\u8c61\u9ad8\u5ea6\u7ed3\u6784\u5316<\/td>\n<td>TCRG\u3001KG\u3001TDRG\u3001AST waveform tracing\u3001structured reports<\/td>\n<td>AlphaChip \u7684 netlist graph\u3001CircuitNet\/ForgeEDA \u7684\u591a\u6a21\u6001\u8868\u793a\u3001OpenDB\/OpenROAD<\/td>\n<td>\u957f\u4e0a\u4e0b\u6587\u53ea\u80fd\u88c5\u4e0b\u66f4\u591a\u6587\u672c&#xff0c;\u7ed3\u6784\u5316\u8868\u793a\u624d\u80fd\u88c5\u4e0b\u8bbe\u8ba1\u56e0\u679c\u5173\u7cfb\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u6570\u636e\u7a00\u7f3a\u4e14\u79c1\u6709<\/td>\n<td>ChipNeMo\u3001CraftRTL\u3001ScaleRTL\u3001ACE-RTL \u7684\u9886\u57df\u6570\u636e\/\u5408\u6210\u6570\u636e\/\u9a8c\u8bc1\u6570\u636e<\/td>\n<td>EDA Corpus\u3001CircuitNet\u3001ForgeEDA\u3001OpenLLM-RTL\u3001NSF workshop \u6570\u636e\u57fa\u7840\u8bbe\u65bd\u5efa\u8bae<\/td>\n<td>\u516c\u5171 benchmark \u4f1a\u63a8\u52a8\u7814\u7a76&#xff0c;\u79c1\u6709 trace \u4f1a\u51b3\u5b9a\u4f01\u4e1a agent \u7684\u5b9e\u9645\u5dee\u8ddd\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u4f18\u5316\u76ee\u6807\u4e0d\u6b62\u4e00\u4e2a<\/td>\n<td>\u529f\u80fd\u6b63\u786e\u3001pass&#064;k\u3001coverage\u3001timing debug\u3001script quality\u3001agent success<\/td>\n<td>PPA\u3001wirelength\u3001congestion\u3001TNS\u3001mask throughput\u3001yield\u3001fab cycle time<\/td>\n<td>AI4EDA \u4e0d\u662f\u5355\u4e00\u6392\u884c\u699c\u95ee\u9898&#xff0c;\u800c\u662f\u591a\u76ee\u6807\u5de5\u7a0b\u63a7\u5236\u95ee\u9898\u3002<\/td>\n<\/tr>\n<tr>\n<td>\u843d\u5730\u4f9d\u8d56\u5de5\u5177\u94fe<\/td>\n<td>JARVIS\u3001Timing Agent\u3001Marco\u3001Trace2Skill \u90fd\u628a\u5de5\u5177\u548c agent \u7ed1\u5b9a<\/td>\n<td>Synopsys\/Cadence\/Siemens \u76f4\u63a5\u628a AI \u653e\u8fdb\u73b0\u6709\u5546\u4e1a\u5de5\u5177&#xff0c;OpenROAD \u63d0\u4f9b\u5f00\u653e\u66ff\u4ee3\u5e95\u5ea7<\/td>\n<td>\u6700\u5148\u89c4\u6a21\u5316\u7684\u4f1a\u662f\u201c\u5de5\u5177\u589e\u5f3a\u578b agent\u201d&#xff0c;\u4e0d\u662f\u8131\u79bb EDA \u5de5\u5177\u7684\u901a\u7528\u804a\u5929\u6a21\u578b\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u7edf\u4e00\u6846\u67b6&#xff1a;AI4EDA \/ AI4Design \u7684\u4e94\u5c42\u80fd\u529b\u6808<\/h3>\n<table>\n<tr>\u5c42\u7ea7\u6838\u5fc3\u5bf9\u8c61\u4ee3\u8868\u5de5\u4f5c\u5224\u65ad\u6807\u51c6<\/tr>\n<tbody>\n<tr>\n<td>1. \u53ef\u6267\u884c\u8bc4\u6d4b\u5c42<\/td>\n<td>testbench\u3001formal\u3001simulator\u3001hidden tests\u3001benchmark harness<\/td>\n<td>VerilogEval\u3001FVEval\u3001CVDP\u3001RTLLM<\/td>\n<td>\u80fd\u5426\u4ece\u6587\u672c\u76f8\u4f3c\u8f6c\u5411\u884c\u4e3a\u6b63\u786e\u3001\u8986\u76d6\u5145\u5206\u548c\u53ef\u590d\u73b0\u3002<\/td>\n<\/tr>\n<tr>\n<td>2. \u9886\u57df\u8bed\u6599\u4e0e\u6a21\u578b\u5c42<\/td>\n<td>RTL\u3001\u811a\u672c\u3001\u6587\u6863\u3001bug\u3001\u65e5\u5fd7\u3001tokenizer\u3001RAG\u3001instruction data<\/td>\n<td>ChipNeMo\u3001ChipAlign\u3001JARVIS\u3001EDA Corpus<\/td>\n<td>\u662f\u5426\u80fd\u628a\u56de\u7b54\u548c\u751f\u6210\u7ed1\u5b9a\u5230\u5de5\u5177\u4e8b\u5b9e\u4e0e\u9879\u76ee\u4e0a\u4e0b\u6587\u3002<\/td>\n<\/tr>\n<tr>\n<td>3. \u5de5\u5177\u53cd\u9988\u4e0e agent \u5c42<\/td>\n<td>compiler\u3001simulator\u3001waveform\u3001STA\u3001formal tool\u3001verifier feedback<\/td>\n<td>RTLFixer\u3001VerilogCoder\u3001Timing Agent\u3001PRO-V-R1\u3001ACE-RTL\u3001Trace2Skill<\/td>\n<td>\u662f\u5426\u5f62\u6210 observe-think-act-patch-verify \u7684\u7a33\u5b9a\u95ed\u73af\u3002<\/td>\n<\/tr>\n<tr>\n<td>4. \u591a\u6a21\u6001\u7269\u7406\u8bbe\u8ba1\u5c42<\/td>\n<td>layout image\u3001congestion map\u3001graph\u3001tabular PPA\u3001netlist\u3001AIG<\/td>\n<td>Multimodal PD Assistant\u3001CircuitNet\u3001ForgeEDA\u3001AiEDA\u3001DREAMPlace<\/td>\n<td>\u662f\u5426\u80fd\u628a\u56fe\u50cf\/\u8868\u683c\/\u56fe\u7ed3\u6784\u8f6c\u4e3a\u53ef\u64cd\u4f5c\u8bbe\u8ba1\u5efa\u8bae\u3002<\/td>\n<\/tr>\n<tr>\n<td>5. \u4ea7\u4e1a\u5316\u5168\u94fe\u8def\u5c42<\/td>\n<td>PPA\u3001TNS\u3001TAT\u3001mask throughput\u3001power\u3001cost\u3001\u5148\u8fdb\u5de5\u827a scaling<\/td>\n<td>AlphaChip\u3001OpenROAD\u3001Synopsys.ai\u3001Cadence Cerebrus\u3001cuLitho<\/td>\n<td>\u662f\u5426\u80fd\u5728\u771f\u5b9e flow \u4e2d\u5e26\u6765\u53ef\u91cf\u5316\u7684\u8bbe\u8ba1\u5468\u671f\u3001\u8d28\u91cf\u3001\u6210\u672c\u6536\u76ca\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u8fb9\u754c\u3001\u98ce\u9669\u4e0e\u4e13\u4e1a\u5224\u65ad<\/h3>\n<p>\u7b2c\u4e00&#xff0c;benchmark \u4e0d\u662f signoff\u3002 VerilogEval\/FVEval\/CVDP \u4ee3\u8868\u8bc4\u6d4b\u8fdb\u6b65&#xff0c;\u4f46\u771f\u5b9e SoC \u8fd8\u6d89\u53ca\u79c1\u6709 IP\u3001EDA \u7248\u672c\u3001PDK\u3001license\u3001\u7ea6\u675f\u6587\u4ef6\u3001CDC\/RDC\u3001UPF\u3001\u4f4e\u529f\u8017\u3001DFT\u3001\u5b89\u5168\u548c\u53ef\u9760\u6027\u3002AI agent \u901a\u8fc7 benchmark \u4e0d\u7b49\u4e8e\u53ef\u76f4\u63a5 tapeout\u3002<\/p>\n<p>\u7b2c\u4e8c&#xff0c;LLM \u4e0d\u5e94\u8131\u79bb\u5de5\u5177 oracle\u3002 \u5bf9\u82af\u7247\u8bbe\u8ba1\u800c\u8a00&#xff0c;\u6b63\u786e\u6027\u5fc5\u987b\u7531 compiler\u3001simulator\u3001STA\u3001formal\u3001layout checker\u3001manufacturing model \u7b49 oracle \u5b9a\u4e49\u3002\u6ca1\u6709 verifier \u7684\u201c\u81ea\u4fe1\u56de\u7b54\u201d\u53cd\u800c\u662f\u98ce\u9669\u3002<\/p>\n<p>\u7b2c\u4e09&#xff0c;\u6570\u636e\u57fa\u7840\u8bbe\u65bd\u662f\u957f\u671f\u58c1\u5792\u3002 \u8bba\u6587\u4e2d\u7684\u5f88\u591a\u63d0\u5347\u6765\u81ea\u9886\u57df\u6570\u636e\u3001\u9519\u8bef\u8f68\u8ff9\u3001\u5de5\u5177\u65e5\u5fd7\u548c\u9690\u5f0f\u5de5\u7a0b\u77e5\u8bc6\u3002\u4f01\u4e1a\u80fd\u5426\u6c89\u6dc0 trace store\u3001skill library\u3001verifier schema \u548c\u5b89\u5168 RAG&#xff0c;\u5c06\u51b3\u5b9a AI4EDA \u662f\u5426\u771f\u6b63\u843d\u5730\u3002<\/p>\n<p>\u7b2c\u56db&#xff0c;\u4ea7\u4e1a\u4ef7\u503c\u6700\u7ec8\u770b PPA\/TAT\/\u6210\u672c\u3002 \u5b66\u672f\u4fa7\u91cd pass&#064;k\u3001functional correctness\u3001benchmark score&#xff1b;\u4ea7\u4e1a\u4fa7\u66f4\u5173\u5fc3 PPA\u3001TNS\u3001ECO \u8f6e\u6b21\u3001\u9a8c\u8bc1\u8986\u76d6\u7387\u3001mask \u4ea7\u80fd\u3001\u80fd\u8017\u548c\u5de5\u7a0b\u5e08\u6548\u7387\u3002\u4e24\u7c7b\u6307\u6807\u5fc5\u987b\u6253\u901a\u3002<\/p>\n<h3>\u53c2\u8003\u6765\u6e90\u4e0e\u539f\u6587\u5165\u53e3<\/h3>\n<p>\u672c\u8282\u5217\u51fa\u7528\u4e8e\u6587\u7ae0\u6846\u67b6\u548c\u5916\u90e8\u5ef6\u4f38\u7684\u4e3b\u8981\u6765\u6e90&#xff1b;NVIDIA 19 \u7bc7\u8bba\u6587\u7684 source \u94fe\u63a5\u5df2\u5728\u5404\u5c0f\u8282\u6807\u9898\u5904\u7ed9\u51fa\u3002<\/p>\n<ul>\n<li>NVIDIA cuLitho<\/li>\n<li>Google DeepMind AlphaChip<\/li>\n<li>Circuit Training<\/li>\n<li>DREAMPlace<\/li>\n<li>OpenROAD<\/li>\n<li>EDA Corpus<\/li>\n<li>CircuitNet<\/li>\n<li>ForgeEDA<\/li>\n<li>Chip-Chat<\/li>\n<li>ChipGPT<\/li>\n<li>RTLLM<\/li>\n<li>Synopsys.ai<\/li>\n<li>Cadence Cerebrus<\/li>\n<li>ML for EDA Survey<\/li>\n<li>LLM for EDA Survey<\/li>\n<li>NSF AI4EDA Workshop Report<\/li>\n<\/ul>\n<h4>\u65b0\u589e\u5916\u90e8\u6765\u6e90\u4e0e\u6700\u65b0\u8865\u5145<\/h4>\n<ul>\n<li>Google DeepMind AlphaChip official blog<\/li>\n<li>DREAMPlace GitHub<\/li>\n<li>OpenROAD official site<\/li>\n<li>EDA Corpus paper<\/li>\n<li>CircuitNet paper<\/li>\n<li>ForgeEDA paper<\/li>\n<li>Chip-Chat paper<\/li>\n<li>ChipGPT paper<\/li>\n<li>RTLLM paper<\/li>\n<li>OpenLLM-RTL paper<\/li>\n<li>RTL-BenchLS paper<\/li>\n<li>Machine Learning for EDA survey<\/li>\n<li>LLM for EDA survey<\/li>\n<li>NSF AI4EDA Workshop Report<\/li>\n<li>Synopsys.ai official page<\/li>\n<li>Cadence Cerebrus official page<\/li>\n<li>Cadence Level-5 ChipStack 2026 release<\/li>\n<li>Siemens Solido Generative &amp; Agentic AI<\/li>\n<li>NVIDIA cuLitho<\/li>\n<li>NVIDIA and TSMC fab AI 2026<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a&#xff1a;\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf \u539f\u6587\u5305\u542b\u5927\u91cf\u8bba\u6587\u622a\u56fe\u548c\u6765\u6e90\u94fe\u63a5\u3002\u5b8c\u6574\u7248\u89c1&#xff1a;https:\/\/ai4eda-ai4design-report-20260624-qcy.netlify.app \u516c\u4f17\u53f7 \u201c\u82af\u7247\u4eba-\u6652 AI \u7b14\u8bb0\u201d<br \/>\n\u672c\u6587\u91c7\u7528\u201c\u4e00\u6761\u7eb5\u8f74\u3001\u4e94\u6761\u6a2a\u8f74\u201d\u7684\u7ed3\u6784&#xff1a;NVIDIA 2023-2026 \u5e74 19 \u7bc7\u8bba\u6587\u4f5c\u4e3a\u8fde\u7eed\u7eb5\u8f74&#xff0c;Google\/DeepMind\u3001DREAMP<\/p>\n","protected":false},"author":2,"featured_media":90118,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[97,656],"topic":[],"class_list":["post-90138","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-97","tag-656"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/90138.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a&#xff1a;\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf \u539f\u6587\u5305\u542b\u5927\u91cf\u8bba\u6587\u622a\u56fe\u548c\u6765\u6e90\u94fe\u63a5\u3002\u5b8c\u6574\u7248\u89c1&#xff1a;https:\/\/ai4eda-ai4design-report-20260624-qcy.netlify.app \u516c\u4f17\u53f7 \u201c\u82af\u7247\u4eba-\u6652 AI \u7b14\u8bb0\u201d \u672c\u6587\u91c7\u7528\u201c\u4e00\u6761\u7eb5\u8f74\u3001\u4e94\u6761\u6a2a\u8f74\u201d\u7684\u7ed3\u6784&#xff1a;NVIDIA 2023-2026 \u5e74 19 \u7bc7\u8bba\u6587\u4f5c\u4e3a\u8fde\u7eed\u7eb5\u8f74&#xff0c;Google\/DeepMind\u3001DREAMP\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/90138.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-04T11:09:54+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110947-6a71c87b39e25.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"28 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/90138.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/90138.html\",\"name\":\"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-04T11:09:54+00:00\",\"dateModified\":\"2026-08-04T11:09:54+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/90138.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/90138.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/90138.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/90138.html","og_locale":"zh_CN","og_type":"article","og_title":"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a&#xff1a;\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf \u539f\u6587\u5305\u542b\u5927\u91cf\u8bba\u6587\u622a\u56fe\u548c\u6765\u6e90\u94fe\u63a5\u3002\u5b8c\u6574\u7248\u89c1&#xff1a;https:\/\/ai4eda-ai4design-report-20260624-qcy.netlify.app \u516c\u4f17\u53f7 \u201c\u82af\u7247\u4eba-\u6652 AI \u7b14\u8bb0\u201d \u672c\u6587\u91c7\u7528\u201c\u4e00\u6761\u7eb5\u8f74\u3001\u4e94\u6761\u6a2a\u8f74\u201d\u7684\u7ed3\u6784&#xff1a;NVIDIA 2023-2026 \u5e74 19 \u7bc7\u8bba\u6587\u4f5c\u4e3a\u8fde\u7eed\u7eb5\u8f74&#xff0c;Google\/DeepMind\u3001DREAMP","og_url":"https:\/\/www.wsisp.com\/helps\/90138.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-04T11:09:54+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260804110947-6a71c87b39e25.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"28 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/90138.html","url":"https:\/\/www.wsisp.com\/helps\/90138.html","name":"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-04T11:09:54+00:00","dateModified":"2026-08-04T11:09:54+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/90138.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/90138.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/90138.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"AI4EDA \/ AI4Design \u5b8c\u6574\u7814\u7a76\u62a5\u544a\uff1a\u5168\u7403\u8bba\u6587\u3001\u9879\u76ee\u3001\u673a\u6784\u4e0e\u4ea7\u4e1a\u8def\u7ebf"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/90138","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=90138"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/90138\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/90118"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=90138"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=90138"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=90138"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=90138"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}