Energy-efficient computing-in-memory architecture for AI processor: device, circuit, architecture perspective

Liang Chang, Chenglong Li, Zhaomin Zhang, Jianbiao Xiao, Qingsong Liu, Zhen Zhu, Weihang Li, Zixuan Zhu, Siqi Yang, Jun Zhou
2021 Science China Information Sciences  
An artificial intelligence (AI) processor is a promising solution for energy-efficient data processing, including health monitoring and image/voice recognition. However, data movements between compute part and memory induce memory wall and power wall challenges to the conventional computing architecture. Recently, the memory-centric architecture has been revised to solve the data movement issue, where the memory is equipped with the compute-capable memory technique, namely, computing-in-memory
more » ... CIM). In this paper, we analyze the requirement of AI algorithms on the data movement and low power requirement of AI processors. In addition, we introduce the story of CIM and implementation methodologies of CIM architecture. Furthermore, we present several novel solutions beyond traditional analog-digital mixed static random-access memory (SRAM)-based CIM architecture. Finally, recent CIM tape-out studies are listed and discussed. Algorithm of CONV layer for b = 0, to B do { for k = 0, to K do {for c = 0, to C do {for y = 0, to Y do {for x = 0, to X do, {for f(y) = 0, to FY do {for f(x) = 0, to FX do {O[b][k][x][y] += I[b][c][x+f(x)][y+f(y)] * W[k][c][f(x)][f(y)] }}}}}}} Weight buffer Input buffer PE array Output buffer Kernel ifmap ofmap Pre-processing 2 Preliminary of computing-in-memory and application-driven memorycentric computing In this section, we introduce the history of the traditional CIM technologies and architecture. Previously, the CIM was proposed to breakthrough the memory limitation by combining memory and computing logic onto the same die, as shown in Figure 2 (a). This solution was failed due to the limitation of CMOS technology and application. Recently, the CIM theory is revived thanks to big data and machine learning. This section discusses energy efficiency requirements demanded by the current machine learning,
doi:10.1007/s11432-021-3234-0 fatcat:np7wtg24rzavbc5fsmammikn3i