AFE-CIM: A Current-Domain Compute-In-Memory Macro for Analog-to-Feature Extraction

Sudarshan Sharma, Wei-Chun Wang, Coleman DeLude, Minah Lee, Nael Mizanur Rahman, Narasimha Vasishta Kidambi, Justin Romberg, Saibal Mukhopadhyay

Abstract: An SRAM-based mixed-signal current-domain Compute-In-Memory (CIM) macro is presented that extracts low-dimensional digital features from high-dimensional analog inputs (Analog Feature Extraction, AFE). A prototype AFECIM chip in 28nm CMOS extracts 16-channel (S-bit) digital features from 128-channel analog input $(8\times$ dimension reduction). The chip consumes 71.25 mW at a throughput of 600 Mega-Sample-Per-Second (MSPS, a sample is 128-analog input) achieving 43.7 TOPS/W peak energy efficiency and 120 $TOPS/mm^{2}$ peak computation density.