FLEX-CIM: A Flexible Kernel Size 1-GHz 181.6-TOPS/W 25.63-TOPS/mm² Analog Compute-in-Memory Macro

Abstract

Compute-in-memory (CIM) is a promising approach for realizing energy-efficient convolutional neural network (CNN) accelerators. Previous CIM works demonstrated a high peak energy efficiency of over 100 TOPS/W, with larger fabrics of 1000 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$+$</tex-math> </inline-formula> channels. Yet, they typically suffer from low utilization for small CNN layers (e.g., <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim$</tex-math> </inline-formula> 9% for ResNet-32). It penalizes their average energy efficiency, throughput density, and effective memory size by the utilization rate. In addition, the analog-to-digital converter (ADC) occupies most of their computing time ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim$</tex-math> </inline-formula> 90%), further hindering the CIM’s throughput. This work presents an FLEX-CIM fabricated under 28-nm CMOS featuring: 1) an analog partial sum (APS) circuit to enable a flexible CIM Kernel size; 2) an overclocked fast multiply–accumulate array (FMA) to boost the throughput; and 3) an adaptive-resolution ADC to enhance the throughput and energy efficiency. The achieved utilization is 99.2% on ResNet-32. Under 4-bit MAC precision, the peak energy efficiency is 181.6 TOPS/W, and the peak throughput density is 25.63 TOPS/mm <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$^{2}$</tex-math> </inline-formula> .