Abstract

This paper presents a 32x32 pseudo-ReRAM-based analog computing-in-memory (CIM) macro in 28nm CMOS. A 4b self-error-correcting word-line (WL) driver reduces the analog compute inaccuracy while minimizing the latency. A stability compensating dummy row maximizes the accumulation length of the multiply-and-accumulate (MAC). The columnsensing dual-phase 6b successive-approximation-register (SAR) analog-to-digital-converter (ADC) maximizes the through-put with minimized pitch. The proposed CIM occupies an active area of 0.0155mm2 and consumes 4.36mW with an average energy efficiency of 25.8TOPS/W. The measured performance achieves the highest normalized throughput with an end-to-end inference accuracy comparable to FP32 with less than a 0.11% drop.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call