具有最优权重划分和资源分配策略的 24.65 TOPS丨W@INT8 混合模拟数字多核 SRAM CIM 宏.pdf

1、1A 24.65 TOPS/WINT8 Hybrid Analog-Digital Multi-core SRAM CIM Macro with Optimal Weight Dividing and Resource Allocation StrategiesYitong Zhou,Wente Yi,Sifan Sun,Wenjia Wang,Jinyu Bai,He Zhang*,Wang Kang*School of Integrated Circuit Science and Engineering,Beihang UniversityOutlineBackgroundProposed

2、 Multi-core hybrid CIM architectureHybrid Weighting SchemeWeight Divide Strategy&Computing Resource AllocationExperiment&ResultsConclusion2OutlineBackgroundProposed Multi-core hybrid CIM architectureHybrid Weighting SchemeWeight Divide Strategy&Computing Resource AllocationExperiment&ResultsConclusi

3、on34Background Analog CIM Analog CIM:Use physical quantities to represent data and completes the analog calculations based on certain physical lawsCurrent-domain computing paradigm Charge-domain computing paradigmTime-domain computing paradigmExtremely high energy efficiency at medium to low precisi

4、onComputational errorsSubstantial area and power overhead of peripherals at medium to high precisionA 351 TOPS/W and 372.4 GOPS Compute-in-Memory SRAM Macro in 7nm FinFET CMOS for Machine-Learning Applications(ISSCC 2020)5Background Digital CIM Digital CIM:Integrate logic gates into memory cells to

5、perform operations in digital domain No calculation loss Significant areaoverhead of peripherals at medium to low precisionA 5-nm 254-TOPS/W 221-TOPS/mm Fully-Digital Computing-in-Memory Macro Supporting Wide-Range Dynamic-Voltage-Frequency Scaling and Simultaneous MAC and Write Operations(2022 ISSC

6、C)6Background Hybrid CIM Performance comparison of ACIM and DCIM with different precisionL2M PrecisionM2H PrecisionACIMDCIM ACIMDCIMEnergyEfficiencyAreaEfficiencyCalculationAccuracyCategoryIndex7Background Related WorkA 28nm 157TOPS/W 446.9Kb/mm2 Compute-In-Memory SRAM Macro with Analog-Digital Hybr