A Switched-Capacitor Integer Compute Unit with Decoupled Storage and Arithmetic for Cloud AI Inference in 5nm CMOS

This paper presents a switched-capacitor-based integer compute unit in 5nm CMOS that is designed as a drop-in replacement for an equivalent digital unit to improve power efficiency by 2. 5X. Integer multiply-accumulate (MAC) operations are recast as a scaled sum of l-b MACs, where each l-b MAC is performed using a population counter (PPCTR) circuit. Each PPCTR is an enhanced SAR ADC that performs l-b multiplication, D-A conversion, accumulation, and A-D conversion with no loss of precision. The compute unit has 4864 PPCTRs arranged as 64 processing engines, with a total throughput of 104.9 TOPS and 650 TOPS/W power efficiency for l-b MACs.