JSSC 2023第7期Memory22nmSRAMNeural Network Accelerator

IMPACT A 1-to-4b 813-TOPSW 22-nm FD-SOI Compute-in-Memory CNN Accelerator Featur

IMPACT是一款22nm FD-SOI工艺的低精度边缘CNN加速器，具有813 TOPS/W的计算效率。

813 TOPS/W, 64 MHz, 4.2 POPS/W, 146 TOPS/mm2

计算内存低精度CNN模拟批量归一化FD-SOI能量效率

▸创新点1：72-kB双电源CIM-SRAM宏（电路创新）：采用6T-based DP操作符，支持1-to-4b混合信号处理，显著提升计算效率和能效，峰值能效达4.2 POPS/W，面积效率达146 TOPS/mm²。

▸创新点2：多比特模拟批量归一化单元（电路创新）：通过模拟信号处理绕过ADC量化问题，减少信息损失，提升CNN性能，尤其在低精度边缘计算场景中表现优异。

▸创新点3：CIM感知的CNN训练框架（方法创新）：针对CIM宏的模拟非线性和变异性进行优化，设计了一套协同训练的CNN框架，显著提升模型在实际硬件上的表现。

▸创新点4：高度并行、通道和精度自适应的数字数据路径（系统创新）：集成CIM-SRAM宏，支持内存传输和输入重塑功能，提升整体加速器的并行处理能力和灵活性。

Abstract

Amid a strife for ever-growing AI processing capa- bilities at the edge, compute-in-memory (CIM) SRAMs involving current-based dot-product (DP) operators have become excel- lent candidates to execute low-precision convolutional neural networks (CNNs) with tremendous energy efficiency. Yet, these architectures suffer from noticeable analog non-idealities and a lack of dynamic range adaptivity, leading to significant information loss during ADC quantization that hinders CNN performance with digita