\u5173\u952e\u8bcd：AI4EDA\u3001Agentic EDA\u3001AlphaChip\u3001RTL \u751f\u6210\u3001\u5f3a\u5316\u5b66\u4e60\u5e03\u5c40\u3001ChipNeMo\u3001VerilogEval\u3001\u5f62\u5f0f\u9a8c\u8bc1\u3001\u9886\u57df\u5fae\u8c03<\/p>\n
\u82af\u7247\u8bbe\u8ba1\u7684\u590d\u6742\u5ea6\u5927\u7ea6\u6bcf\u4e24\u5e74\u7ffb\u4e00\u500d，\u800c\u5de5\u7a0b\u5e08\u751f\u4ea7\u529b\u53ea\u4ee5\u7ea6 1.2\u00d7 \u7684\u901f\u5ea6\u589e\u957f\u2014\u2014\u8fd9\u9053\u201c\u751f\u4ea7\u529b\u9e3f\u6c9f\u201d\u6b63\u662f AI \u6d8c\u5165 EDA（\u7535\u5b50\u8bbe\u8ba1\u81ea\u52a8\u5316）\u7684\u7ed3\u6784\u6027\u52a8\u56e0\u3002\u672c\u6587\u68b3\u7406 AI4EDA \u7684\u4e09\u7ea7\u8dc3\u8fc1\u6846\u67b6（L1 \u4f20\u7edf\u5de5\u5177 \u2192 L2 AI for EDA \u2192 L3 Agentic EDA \u2192 L4 \u81ea\u4e3b\u8bbe\u8ba1），\u62c6\u89e3 AlphaChip \u7528\u5f3a\u5316\u5b66\u4e60\u505a TPU \u5e03\u5c40\u3001AlphaEvolve \u7528\u201cLLM \u5019\u9009 + \u81ea\u52a8\u8bc4\u6d4b + \u8fdb\u5316\u641c\u7d22\u201d\u91cd\u5199 Verilog \u7b97\u672f\u7535\u8def\u3001ChipNeMo \u505a\u9886\u57df\u5fae\u8c03\u3001\u4ee5\u53ca LLM \u751f\u6210 \/ \u9a8c\u8bc1 RTL（VeriGen\u3001AutoChip\u3001Spec2RTL-Agent\u3001VeriReason\u3001QiMeng-CodeV-R1，VerilogEval pass@1 \u5df2\u8fbe 65%~70%）\u7684\u6838\u5fc3\u539f\u7406；\u7ed9\u51fa\u201c\u89c4\u683c \u2192 RTL \u2192 \u9a8c\u8bc1 \u2192 \u4fee\u6b63\u201d\u95ed\u73af\u7684\u5b9e\u64cd\u6b65\u9aa4，\u5e76\u9644\u5e26\u4e00\u4e2a\u7528 NumPy \u5b9e\u73b0\u7684 RL \/ \u542f\u53d1\u5f0f\u5b8f\u5e03\u5c40（HPWL \u6700\u5c0f\u5316）\u53ef\u8fd0\u884c Demo（\u5df2\u5b9e\u6d4b：\u76f8\u5bf9\u968f\u673a\u5e03\u5c40\u7ebf\u957f\u6539\u5584 46.7%）\u3002\u6700\u540e\u603b\u7ed3 12 \u6761 FAQ \u4e0e 7 \u9879\u4f18\u5316\u7b56\u7565，\u5e76\u5398\u6e05\u201cAI \u80fd\u8bbe\u8ba1\u82af\u7247\u201d\u4e0e\u201cAI \u5728\u90e8\u5206\u73af\u8282\u8d85\u8d8a\u4eba\u7c7b\u201d\u7684\u8fb9\u754c\u3002<\/p>\n
\u534a\u5bfc\u4f53\u8bbe\u8ba1\u590d\u6742\u5ea6（\u4ee5\u6676\u4f53\u7ba1\u6570\u8ba1）\u7ea6\u6bcf 2 \u5e74\u7ffb\u500d，\u800c EDA \u5de5\u5177\u5e26\u6765\u7684\u8bbe\u8ba1\u5e08\u751f\u4ea7\u529b\u63d0\u5347\u4ec5\u7ea6 1.2\u00d7 \/ 2 \u5e74\u3002\u4e00\u4e2a 2nm SoC \u5305\u542b\u6570\u767e\u4ebf\u6676\u4f53\u7ba1，RTL-to-GDSII（\u5230\u5236\u9020\u63a9\u819c\u6570\u636e）\u6d41\u7a0b\u8dd1\u51e0\u767e\u6b65，\u65f6\u5e8f \/ \u529f\u8017 \/ DRC \u7ea6\u675f\u8d85\u8fc7\u4e00\u4e07\u6761\u3002\u7f29\u653e\u4eba\u7c7b\u56e2\u961f\u662f\u7ebf\u6027\u7684，\u590d\u6742\u5ea6\u7f29\u653e\u662f\u6307\u6570\u7684\u2014\u2014\u8fd9\u9053\u9e3f\u6c9f\u53ea\u80fd\u9760\u81ea\u52a8\u5316\u586b\u8865\u3002<\/p>\n
\u82af\u7247\u8bbe\u8ba1\u6570\u636e\u4e09\u7279\u5f81：\u91cf\u5c11（\u76f8\u5bf9\u4e92\u8054\u7f51\u6587\u672c）\u3001\u79c1\u6709（IP \u4fdd\u62a4\u4e0e\u5546\u4e1a\u673a\u5bc6）\u3001\u5f02\u6784（RTL\u3001\u7ea6\u675f\u6587\u4ef6\u3001\u65f6\u5e8f\u62a5\u544a\u3001\u6ce2\u5f62\u3001\u5e03\u5c40\u56fe，\u683c\u5f0f\u5b8c\u5168\u4e0d\u7edf\u4e00）\u3002\u4e0d\u80fd\u50cf\u8bad\u7ec3\u901a\u7528\u5927\u6a21\u578b\u90a3\u6837\u722c\u4e92\u8054\u7f51，\u53ea\u80fd\u201c\u81ea\u4ea7\u81ea\u9500\u201d\u6216\u4ece\u5de5\u5177\u8fd0\u884c\u8f68\u8ff9\u91cc\u69a8\u4fe1\u53f7\u3002<\/p>\n
\u4e1a\u754c\u628a\u82af\u7247\u8bbe\u8ba1 AI \u7684\u6f14\u8fdb\u5206\u6210 L1~L4（\u6e90\u81ea 2025 \u5e74\u5e95\u53d1\u5e03\u7684 Agentic EDA \u7efc\u8ff0 arXiv:2512.23189）：<\/p>\n
\u6211\u4eec\u6b63\u5904\u5728 L2 \u2192 L3 \u7684\u8fc7\u6e21\u671f\u3002<\/p>\n
DeepMind 2020 \u5e74\u5728 Nature \u53d1\u8868\u7684\u65b9\u6cd5，\u7528\u5f3a\u5316\u5b66\u4e60\u81ea\u52a8\u751f\u6210\u82af\u7247\u5e03\u5c40，\u5728\u201c\u7ebf\u957f + \u65f6\u5e8f\u201d\u590d\u5408\u76ee\u6807\u4e0a\u8d85\u8d8a\u4eba\u7c7b\u4e13\u5bb6，\u4e14\u80fd\u8fc1\u79fb\u5230\u672a\u89c1\u8fc7\u7684\u8bbe\u8ba1\u3002\u5b83\u5df2\u88ab\u7528\u4e8e Google TPU v5 \u53ca\u4e4b\u540e\u4e09\u4ee3（\u4ee5\u53ca\u6570\u636e\u4e2d\u5fc3 CPU Axion），\u662f\u201c\u771f\u7845\u7247\u3001\u975e\u7eaf\u8bba\u6587\u201d\u3002\u4f46\u5b83\u7684\u8986\u76d6\u53ea\u6709\u6d41\u7a0b 9 \u6b65\u91cc\u7684\u7ea6 4 \u6b65（\u5e03\u5c40\u3001P&R\u3001\u65f6\u5e8f\u9884\u6d4b\u3001OPC），RTL \u8bbe\u8ba1\u3001\u9a8c\u8bc1\u3001\u6d4b\u8bd5\u8bbe\u8ba1\u5b83\u90fd\u4e0d\u78b0\u3002\u6240\u4ee5\u51c6\u786e\u8bf4\u6cd5\u662f\u201cAI \u5728\u82af\u7247\u8bbe\u8ba1\u7684\u90e8\u5206\u73af\u8282\u8d85\u8d8a\u4eba\u7c7b\u201d，\u800c\u975e\u201cAI \u80fd\u8bbe\u8ba1\u82af\u7247\u201d\u3002<\/p>\n
AlphaEvolve \u4ee3\u8868\u66f4\u5951\u5408 EDA \u7684\u8def\u7ebf：LLM \u4ea7\u751f\u4ee3\u7801\u5019\u9009 \u2192 \u81ea\u52a8\u8bc4\u6d4b\u5668\u8d1f\u8d23\u6b63\u786e\u6027\u4e0e\u6027\u80fd\u6253\u5206 \u2192 \u5019\u9009\u5728\u8fdb\u5316\u5faa\u73af\u4e2d\u88ab\u9009\u62e9\u3001\u53d8\u5f02\u3001\u4fdd\u7559\u3002DeepMind \u66fe\u7528\u5b83\u63d0\u51fa\u9488\u5bf9 TPU \u77e9\u9635\u4e58\u6cd5\u7b97\u672f\u7535\u8def\u7684 Verilog \u6539\u5199，\u5220\u6389\u4e0d\u5fc5\u8981\u7684 bit，\u9a8c\u8bc1\u901a\u8fc7\u540e\u96c6\u6210\u8fdb upcoming TPU；2026 \u5e74\u66f4\u65b0\u79f0\u5df2\u6210\u4e3a\u4f18\u5316\u4e0b\u4e00\u4ee3 TPU \u7684\u5e38\u89c4\u5de5\u5177\u3002\u542f\u793a：\u5bf9\u53ef\u8bc4\u5206\u3001\u53ef\u4eff\u771f\u3001\u53ef\u641c\u7d22\u7684\u95ee\u9898，RL \/ \u8fdb\u5316\u641c\u7d22 \/ \u8d1d\u53f6\u65af\u4f18\u5316 \/ EDA \u542f\u53d1\u5f0f\u4ecd\u6bd4\u901a\u7528 LLM \u66f4\u9002\u5408；LLM \u7684\u5f3a\u9879\u662f\u8bfb\u89c4\u683c\u3001\u63d0\u5019\u9009\u3001\u89e3\u91ca\u65e5\u5fd7\u3001\u7ec4\u7ec7\u641c\u7d22\u3001\u5199\u80f6\u6c34\u4ee3\u7801\u3002<\/p>\n
NVIDIA \u7528 custom tokenizer + \u9886\u57df\u81ea\u9002\u5e94\u9884\u8bad\u7ec3 + \u9886\u57df\u6307\u4ee4\u5bf9\u9f50，\u5728 assistant\u3001\u811a\u672c\u751f\u6210\u3001bug \u6458\u8981\u7b49\u4efb\u52a1\u4e0a\u9a8c\u8bc1\u4e86\u9886\u57df\u9002\u914d\u6536\u76ca\u3002\u7ed3\u8bba\u5f88\u6e05\u695a：\u82af\u7247\u8bbe\u8ba1\u7684\u8bed\u8a00（RTL\u3001netlist\u3001EDA \u65e5\u5fd7\u3001\u65f6\u5e8f\u62a5\u544a）\u8bcd\u6cd5\u548c\u4e0a\u4e0b\u6587\u90fd\u7279\u6b8a，\u901a\u7528\u5927\u6a21\u578b\u76f4\u63a5\u4e0a\u6548\u679c\u6709\u9650，\u5fc5\u987b\u9886\u57df\u5fae\u8c03\u3002<\/p>\n
\u4ece VeriGen\u3001AutoChip \u6f14\u8fdb\u5230 Spec2RTL-Agent\u3001VeriReason\u3001QiMeng \u7cfb\u5217（CodeV-R1 \u7b49），\u7ed3\u5408 RAG \/ RLVR \/ \u5f62\u5f0f\u9a8c\u8bc1\u53cd\u9988\u505a\u8fed\u4ee3\u4fee\u6b63，\u5728 VerilogEval \u57fa\u51c6 pass@1 \u5df2\u7a81\u7834 65%~70%\u3002\u8fd9\u610f\u5473\u7740\u5bf9\u5927\u91cf\u5e38\u89c4\u6a21\u5757，LLM \u5df2\u80fd\u4ea7\u51fa\u201c\u5927\u6982\u7387\u53ef\u7efc\u5408\u3001\u5927\u6982\u7387\u6b63\u786e\u201d\u7684 RTL，\u518d\u4ea4\u7531\u81ea\u52a8\u9a8c\u8bc1\u95ed\u73af\u6253\u78e8\u3002<\/p>\n
\u5f00\u6e90\u9879\u76ee\u964d\u4f4e\u4e86\u95e8\u69db：Croc SoC\u3001CHIA\u3001OpenROAD Agent\u3001ASIC-Agent；\u5f00\u653e\u6570\u636e\u6709 OpenLLM-RTL\u3001EDA Corpus\u3001CircuitNet\u3001ForgeEDA\u3002NSF AI4EDA Workshop Report \u76f4\u63a5\u628a\u201c\u6570\u636e\u57fa\u7840\u8bbe\u65bd\u201d\u5217\u4e3a\u5b66\u79d1\u7ea7\u4f18\u5148\u4e8b\u9879\u2014\u2014\u6ca1\u6709\u516c\u5171 benchmark，\u9886\u57df\u5c31\u65e0\u6cd5\u5f62\u6210\u516c\u5171\u8fdb\u6b65\u3002<\/p>\n
NVIDIA 2026 \u5e74\u8bba\u6587\u7ed9\u51fa Agentic EDA \u7684\u8868\u793a\u65b9\u6848：TCRG（Tapeout Context Representation Graph） \u628a\u8bbe\u8ba1\u4e0a\u4e0b\u6587\u7f16\u7801\u4e3a\u56fe\u3001Knowledge Graph \u628a\u8bbe\u8ba1\u77e5\u8bc6\u7ed3\u6784\u5316\u3001Timing Dependency Rule Graph \u628a\u65f6\u5e8f\u7ea6\u675f\u8868\u793a\u4e3a\u89c4\u5219\u56fe\u3001AST Waveform Tracing \u628a\u6ce2\u5f62\u8f6c\u6210\u62bd\u8c61\u8bed\u6cd5\u6811\u7ea7\u7ed3\u6784\u5316\u8ffd\u8e2a\u3002\u6838\u5fc3\u89c2\u70b9：\u957f\u4e0a\u4e0b\u6587\u662f\u7f13\u5175\u4e4b\u8ba1，\u7ed3\u6784\u5316\u8868\u793a\u624d\u662f\u6839\u672c\u89e3\u2014\u2014\u628a netlist\u3001\u7ea6\u675f\u3001\u65f6\u5e8f\u62a5\u544a\u3001\u6ce2\u5f62\u8f6c\u5316\u4e3a Agent \u53ef\u76f4\u63a5\u63a8\u7406\u7684\u7ed3\u6784\u5316\u4e2d\u95f4\u8868\u793a\u3002\u8fd9\u4e5f\u662f\u77e5\u8bc6\u56fe\u8c31\u5728 AI4EDA \u4e2d\u7684\u4ef7\u503c\u8fdc\u9ad8\u4e8e\u901a\u7528 NLP \u7684\u539f\u56e0：\u82af\u7247\u8bbe\u8ba1\u5929\u7136\u5c31\u662f\u56fe\u7ed3\u6784\u3002<\/p>\n
\u4e0b\u9762\u662f\u4e00\u5957\u201c\u89c4\u683c \u2192 RTL \u2192 \u9a8c\u8bc1 \u2192 \u4fee\u6b63\u201d\u7684 L3 \u95ed\u73af\u843d\u5730\u53c2\u8003\u3002<\/p>\n
\u6b65\u9aa4 1：\u660e\u786e\u81ea\u6cbb\u7b49\u7ea7\u4e0e\u73af\u8282\u3002 \u5148\u51b3\u5b9a\u54ea\u4e9b\u73af\u8282\u5148\u4e0a Agent（\u901a\u5e38 RTL \u751f\u6210\u3001\u811a\u672c\u751f\u6210\u3001bug \u6458\u8981\u3001\u7269\u7406\u8bbe\u8ba1\u8f85\u52a9\u4f18\u5148；\u7b7e\u6838\u3001\u6a21\u62df\u4ecd\u4eba\u673a\u6df7\u5408）\u3002<\/p>\n
\u6b65\u9aa4 2：\u6784\u5efa\u9886\u57df\u6570\u636e\u96c6\u3002 \u590d\u7528 ChipNeMo \/ CraftRTL \/ ScaleRTL \/ ACE-RTL \u601d\u8def：\u4ece\u5de5\u5177\u8fd0\u884c\u8f68\u8ff9（\u9a8c\u8bc1\u5931\u8d25\u65e5\u5fd7\u3001\u7efc\u5408\u62a5\u544a）\u81ea\u4ea7\u81ea\u9500，\u6216\u63a5\u5165 OpenLLM-RTL\u3001EDA Corpus \u7b49\u516c\u5f00\u6570\u636e\u3002<\/p>\n
\u6b65\u9aa4 3：\u9886\u57df\u5fae\u8c03 \/ \u5bf9\u9f50\u3002 \u7528 custom tokenizer + \u9886\u57df\u81ea\u9002\u5e94\u9884\u8bad\u7ec3 + \u6307\u4ee4\u5bf9\u9f50，\u8ba9\u6a21\u578b\u7406\u89e3 RTL \/ \u65f6\u5e8f\u62a5\u544a\u7684\u7279\u6b8a\u8bcd\u6cd5\u3002<\/p>\n
\u6b65\u9aa4 4：RTL \u751f\u6210 + \u81ea\u52a8\u9a8c\u8bc1\u95ed\u73af\u3002 \u751f\u6210 Verilog \u540e\u63a5 iverilog \/ cocotb \/ \u5f62\u5f0f\u9a8c\u8bc1，\u628a\u5931\u8d25\u8f68\u8ff9\u5582\u56de\u6a21\u578b\u505a RLVR \u5f0f\u4fee\u6b63（\u89c1\u7b2c\u56db\u8282\u4ee3\u7801）\u3002<\/p>\n
\u6b65\u9aa4 5：\u7269\u7406\u8bbe\u8ba1 AI \u8f85\u52a9\u3002 \u5e03\u5c40 \/ \u65f6\u5e8f\u9884\u6d4b\u7528 RL \/ GNN，\u4e14\u5f15\u5165\u7269\u7406\u4fe1\u606f\u635f\u5931\u4fdd\u8bc1 DRC \/ LVS \/ PPA \u53ef\u884c\u3002<\/p>\n
\u6b65\u9aa4 6：\u4eba\u673a\u6df7\u5408\u7b7e\u6838\u3002 \u7edd\u5bf9\u786e\u5b9a\u6027\u73af\u8282\u4ecd\u7531\u4eba + \u89c4\u5219\u5de5\u5177\u628a\u5173，AI \u8d1f\u8d23\u63a2\u7d22\u4e0e\u5e38\u89c4\u4efb\u52a1\u3002<\/p>\n
\u4e0b\u9762\u811a\u672c（\u968f\u6587\u9644 _ai4eda_demo.py）\u7528 2D \u7f51\u683c\u73a9\u5177\u95ee\u9898\u590d\u73b0 AlphaChip \u7c7b\u65b9\u6cd5\u7684\u6838\u5fc3\u601d\u60f3：\u7528\u53ef\u8bc4\u5206\u76ee\u6807（\u534a\u5468\u957f\u7ebf\u957f HPWL）\u9a71\u52a8\u641c\u7d22\u3002<\/p>\n
import<\/span> numpy as<\/span> np<\/p>\n def<\/span> hpwl<\/span>(<\/span>positions,<\/span> nets)<\/span>:<\/span> def<\/span> hill_climb<\/span>(<\/span>positions,<\/span> nets,<\/span> grid,<\/span> steps,<\/span> rng)<\/span>:<\/span> # \u5b9e\u6d4b：\u7f51\u683c 8×8 \/ 6 \u5b8f \/ 12 \u7ebf\u7f51，\u968f\u673a\u6700\u597d HPWL=45.0，\u722c\u5c71\u6700\u7ec8 24.0，\u7ebf\u957f\u6539\u5584 46.7%<\/span><\/p>\n \u53c2\u6570\u89e3\u8bfb\u4e0e\u5b9e\u6d4b\u7ed3\u8bba：<\/p>\n # \u4f2a\u4ee3\u7801：Spec -> RTL -> \u5f62\u5f0f\/\u4eff\u771f\u9a8c\u8bc1 -> RLVR \u4fee\u6b63 \u95ed\u73af<\/span>
\n """\u534a\u5468\u957f\u7ebf\u957f：\u5e03\u5c40\u9636\u6bb5\u6700\u5e38\u7528\u7684\u53ef\u5fae\u5206\u4ee3\u7406\u76ee\u6807，\u662f\u771f\u5b9e\u7ebf\u957f\u7684\u4e0b\u754c\u3002"""<\/span>
\n total =<\/span> 0.0<\/span>
\n for<\/span> net in<\/span> nets:<\/span>
\n xs =<\/span> positions[<\/span>net,<\/span> 0<\/span>]<\/span>;<\/span> ys =<\/span> positions[<\/span>net,<\/span> 1<\/span>]<\/span>
\n total +=<\/span> (<\/span>xs.<\/span>max<\/span>(<\/span>)<\/span> –<\/span> xs.<\/span>min<\/span>(<\/span>)<\/span>)<\/span> +<\/span> (<\/span>ys.<\/span>max<\/span>(<\/span>)<\/span> –<\/span> ys.<\/span>min<\/span>(<\/span>)<\/span>)<\/span>
\n return<\/span> float<\/span>(<\/span>total)<\/span><\/p>\n
\n n_macros =<\/span> positions.<\/span>shape[<\/span>0<\/span>]<\/span> # \u5b8f\u6a21\u5757\u6570\u91cf\u4ece\u5e03\u5c40\u5f20\u91cf\u63a8\u5bfc<\/span>
\n best =<\/span> positions.<\/span>copy(<\/span>)<\/span>
\n best_cost =<\/span> hpwl(<\/span>best,<\/span> nets)<\/span>
\n occupied =<\/span> set<\/span>(<\/span>map<\/span>(<\/span>tuple<\/span>,<\/span> positions.<\/span>astype(<\/span>int<\/span>)<\/span>)<\/span>)<\/span>
\n for<\/span> _ in<\/span> range<\/span>(<\/span>steps)<\/span>:<\/span>
\n i =<\/span> rng.<\/span>integers(<\/span>n_macros)<\/span>
\n old =<\/span> best[<\/span>i]<\/span>.<\/span>copy(<\/span>)<\/span>
\n while<\/span> True<\/span>:<\/span> # \u968f\u673a\u6311\u4e00\u4e2a\u7a7a\u95f2\u683c<\/span>
\n cand =<\/span> (<\/span>rng.<\/span>integers(<\/span>grid)<\/span>,<\/span> rng.<\/span>integers(<\/span>grid)<\/span>)<\/span>
\n if<\/span> cand not<\/span> in<\/span> occupied or<\/span> cand ==<\/span> tuple<\/span>(<\/span>old.<\/span>astype(<\/span>int<\/span>)<\/span>)<\/span>:<\/span>
\n break<\/span>
\n best[<\/span>i]<\/span> =<\/span> np.<\/span>array(<\/span>cand,<\/span> dtype=<\/span>float<\/span>)<\/span>
\n occupied.<\/span>discard(<\/span>tuple<\/span>(<\/span>old.<\/span>astype(<\/span>int<\/span>)<\/span>)<\/span>)<\/span>;<\/span> occupied.<\/span>add(<\/span>cand)<\/span>
\n new_cost =<\/span> hpwl(<\/span>best,<\/span> nets)<\/span>
\n if<\/span> new_cost <<\/span> best_cost:<\/span> # \u4e00\u9636\u8d2a\u5fc3：\u53ea\u63a5\u53d7\u964d\u4f4e HPWL \u7684\u79fb\u52a8<\/span>
\n best_cost =<\/span> new_cost
\n else<\/span>:<\/span> # \u4e0d\u63a5\u53d7\u5219\u56de\u9000<\/span>
\n occupied.<\/span>discard(<\/span>cand)<\/span>;<\/span> occupied.<\/span>add(<\/span>tuple<\/span>(<\/span>old.<\/span>astype(<\/span>int<\/span>)<\/span>)<\/span>)<\/span>;<\/span> best[<\/span>i]<\/span> =<\/span> old
\n return<\/span> best,<\/span> best_cost<\/p>\n\n
4.2 RTL \u751f\u6210 + \u9a8c\u8bc1\u95ed\u73af（\u4ee3\u7801\u7247\u6bb5，\u793a\u610f\u53ef\u96c6\u6210\u8fdb\u6d41\u6c34\u7ebf）<\/h4>\n
\ndef<\/span> rtl_loop<\/span>(<\/span>spec,<\/span> model,<\/span> verifier,<\/span> max_iter=<\/span>5<\/span>)<\/span>:<\/span>
\n for<\/span> it in<\/span> range<\/span>(<\/span>max_iter)<\/span>:<\/span>
\n rtl =<\/span> model.<\/span>generate(<\/span>f"\u6839\u636e\u4ee5\u4e0b\u89c4\u683c\u751f\u6210\u53ef\u7efc\u5408 Verilog:\\\\n<\/span>{<\/span>spec}<\/span><\/span>"<\/span><\/span>)<\/span>
\n result =<\/span> verifier.<\/span>run(<\/span>rtl)<\/span> # iverilog \/ cocotb \/ \u5f62\u5f0f\u9a8c\u8bc1<\/span>
\n if<\/span> result.<\/span>pass_:<\/span>
\n return<\/span> rtl # \u901a\u8fc7\u5373\u4ea4\u4ed8<\/span>
\n # \u628a\u5931\u8d25\u8f68\u8ff9\u4f5c\u4e3a RLVR \/ \u53cd\u9988\u5582\u56de，\u4e0b\u4e00\u8f6e\u4fee\u6b63<\/span>
\n model.<\/span>feedback(<\/span>f"\u9a8c\u8bc1\u5931\u8d25：<\/span>{<\/span>result.<\/span>trace}<\/span><\/span>\\\\n\u8bf7\u4fee\u6b63\u4e0a\u8ff0 RTL\u3002"<\/span><\/span>