A 28-nm Static-Power-Free Fully Parallel RRAM-Based TD CIM Macro With 1982 TOPS/W/Bit for Edge Applications
Song-Tao Wei, Peng Yao, Xinying Guo, Dong Wu, Lu Jie, Qi Qin, Bin Gao, Jianshi Tang, He Qian, Sining Pan, Huaqiang Wu
Abstract
Analog computing in memory (CIM) based on resistive nonvolatile memory (NVM) has encountered several issues, such as low parallelism, low computing accuracy, and considerable power consumption. In this letter, a temporal unit based on design technology co-optimization (DTCO) for resistive random access memory is proposed for the first time, with the advantage of eliminating dc current and reducing the deviation of mapped weight. A time-domain (TD) array based on the proposed temporal unit features performing fully parallel matrix-vector multiplication (MVM) in a static-power-free manner, without the consideration of IR drop and limited sensing margin (SM). Besides, a low-power time-digital converter (TDC) with local offset elimination further boosts energy efficiency (EF) and computing accuracy. The fabricated 28-nm TD CIM macro achieves a state-of-the-art normalized EF of 1982 and 1387 TOPS/W/bit under 1b-input, ternary-weight and 4b-input, signed 4b-weight, respectively.