Energy optimization framework for eDRAM based PIM with reconfigurable voltage swing and retention-aware scheduling

Abstract

Processing-in-memory (PIM) has demonstrated its high usage in accelerating data intensive applications such as transformer. Previous works mainly focus on optimizing the computational part of PIM to achieve higher energy efficiency. However, the importance of memory design, which consumes the most power in PIM operation, has been rather neglected. In this work, we propose an energy optimization framework for eDRAM based PIM called RED, featuring a novel reconfigurable eDRAM and retention-aware scheduling. The RED framework dynamically adjusts eDRAM operations to match specific use cases, significantly improving power efficiency. It estimates overall energy consumption across all possible tiling schemes and memory operations, pinpointing the most energy-efficient operating point. Reconfigurable eDRAM demonstrates a reduction in power consumption of memory access by up to 71.31%. Additionally, by selecting the best tiling scheme and memory operation through our energy modeling, RED achieves 1.84×-3.03× higher energy efficiency compared to the baseline.