节能 AI 硬件：神经形态电路和工具.pdf

《节能 AI 硬件：神经形态电路和工具.pdf》由会员分享，可在线阅读，更多相关《节能 AI 硬件：神经形态电路和工具.pdf（14页珍藏版）》请在三个皮匠报告上搜索。

1、OCP Regional Summit|April 19,2023|Prague,CZEnergy Efficient AI Hardware:Neuromorphic Circuits and ToolsDr.Loreto MateuFraunhofer Institute for Integrated Circuits IIS MOTIVATION:FROM THE CLOUD TO THE EDGEWHAT ARE THE ADVANTAGES OF EDGE AI?Low latencyHigh energy efficiencyPrivacyWHAT ARE THE OBJECTIV

2、ES?Ultra low energy consumptionper InferenceUltra-low processingtime per inferenceNeuromorphic ASICs in establishedqualified semiconductorprocessesConfigurable and scalable multi-corearchitecturefor adapted processingSmall area for low priceArchitecturesANNs:Multi-core with NoC e.g.analog,mixed-sign

3、al,digitalSNNs:Multi-core with mesh routing e.g.mixed-signalCircuits:integrated circuits for synapses,neurons,etc.Algorithms:optimization based on the use caseSoftware tools:Hardware-aware training,mapper&compiler,etc.Physical devices Embedded non-volatile memories 6 ENABLERS FOR NEUROMORPHIC COMPUT

4、ING ArchitecturesCircuitsAlgorithmsSoftware ToolsPhysical deviceseNVMsHigh parallelism High computing speedEnergy efficiencyANALOG IN-MEMORY COMPUTING=+x1x2x3xny1y2y3 ymw11Accurate NN computation is necessary even with PVT variations and mismatch of weightsANALOG IN-MEMORY COMPUTING=+HARDWARE-AWARE

5、TRAININGFault-Aware Training(FAT)Model Export Robust NNsModel ExchangeQuantization-Aware Training(QAT)Low Memory FootprintIn an inference accelerator with a multi-core architecture,the layers are mapped into coresMapping for In-Memory Computing can follow different strategies Reduction of data movem

6、entMaximum utilization of hardware resourcesThe mapper has a huge impact on KPIs likeThroughput Energy consumptionMAPPING OF NNs ON THE HARDWAREVAD Network mappingA benchmarking framework is necessary Automated Search is necessary to optimize KPIsEach hardware