Abstract
Resistive random access memory (ReRAM) is a promising technology for AI Processing-in-Memory (PIM) hardware because of its compatibility with CMOS, small footprint, and ability to complete matrix–vector multiplication workloads inside the memory device itself. However, redundant computations are brought on by duplicate weights and inputs when an MVM has to be split into smaller-granularity sequential sub-works in the real world. Recent studies have proposed repetition-pruning to address this issue, but the buffer allocation strategy for enhancing buffer device utilization remains understudied. In preliminary experiments observing input patterns of neural layers with different datasets, the similarity of repetition allows us to transfer the buffer allocation strategy obtained from a small dataset to the computation with a large dataset. Hence, this paper proposes a practical compute-reuse mechanism for ReRAM-based PIM, called CRPIM, which replaces repetitive computations with buffering and reading. Moreover, the subsequent buffer allocation problem is resolved at both inter-layer and intra-layer levels. Our experimental results demonstrate that CRPIM significantly reduces ReRAM cells and execution time while maintaining adequate buffer and energy overhead.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have