芯片设计人工智能导论.pdf

1、Introduction to AI for Chip DesignHaoxing(Mark)Ren,Director of Design Automation Research,NVIDIA08/25/2024AI for Design Performance and ProductivityAnalysisFasterPredictiveCross-StageOptimizationFasterMore scalableBetter resultsAssistanceKnow-howCodingTask automation2AIforChip Design Research NVIDIA

2、2024HPGCN(Testability)20192020202120222023PRIMAL(Power)FIST(PD)PowerNet(IR Drop)ParaGraph(Parasitics)GRANNITE(Switching Activity)NVCell-RL(Cell)MAVIREC(IR Drop)ParaSize(Analog)DOINN(Lithography)Transsizer(PD)TAG(Analog)BufFormer(PD)ILILT(Lithography)ChipNeMo(Engineering)AutoCRAFT RL(Analog)Dream-GAN

3、(PD)AutoDMP(PD)VerilogEval(RTL)RTLFixer(RTL)Clustering(Cell)ClusteringAgent(Cell)VerilogCoder(RTL)DesignAnalysisDesignOptimization(RL,BO)DesignOptimization(Gen AI)DesignAssistanceGraph Cluster(PD)FVAgent(FV)FVEval(FV)OPCAgent(Lithography)We build AI to build chips for AI!PrefixRL(Synthesis)VAESA(Arc

4、h)3 Analysis Classical ML Deep learning Optimization Bayesian optimization Reinforcement learning Optimization Generative AI Assistance LLMAITechniques4AI Techniques Analysis Classical ML Deep learning Optimization Bayesian optimization Reinforcement learning Optimization Generative AI Assistance LL

5、MLinear RegressionSupport Vector MachineDecision TreeNeural Network5Suitable for small structured dataAI Techniques Analysis Classical ML Deep learning Optimization Blackbox optimization Reinforcement learning Optimization Generative AI Assistance LLM925467318CNNSuitable for physical design dataGNNS

6、uitable for circuit netlist data6Faster Analysis IR Drop EstimationIR drop estimation is important for physical design,but it takes hoursUse AI to predict IR drop from cell level features94%accuracy in 3 second vs 3 hr in commercial toolsPower mapsCoefficient maps=1+2+3+4+5+67V.A.Chhabria et al,MAVI