Pro-Cache-CIM: A 28nm 69.4TOPS/W Product-Cache-based Digital-Compute-in-Memory Macro Leveraging Data Locality Pattern in Vision AI Tasks
Wenbin Jia, Yifan He, Xiang Li, Yixuan Xie, Zongle Huanq, Wenxun Wang, B.-S. Chen, Yaolei Li, Jinshan Yue, Xueqing Li, Huazhong Yang, Hongyang Jia, Yongpan Liu · 2025
Digital-based compute-in-memory macros (DCIM) and digital accelerators that leverage lookup tables (LUT) [1, 2, 3] have demonstrated remarkable advantages in achieving high energy efficiency while maintaining full precision. By pre-computing summation results, LUT-based DCIMs eliminate the energy consumption of one or two stages of addertrees. However, scaling these LUT-based DCIMs to higher orders-which requires using wider bit-widths for indexing LUT content-poses significant challenges. This is primarily due to the exponential growth in the combinations of pre-computed results, which consequently leads to proportional increases in area and power consumption, ultimately degrading both area and energy efficiency.