Presentation Information

[17p-M_374-1]Non-votatile memory based Computation-in-Memory for Edge AI

〇Ken Takeuchi1 (1.Univ. Tokyo)

Keywords:

Computation-in-Memory,Non-votatile memory,Edge AI

This paper overviews Computation-in-Memory (CiM) based AI accelerators and neuromorphic computing for edge applications. CiM which uses non-volatile memories such as Re-RAM, FeFET, PRAM, MRAM and Flash memories achieves massively parallel MAC (Multiply-Accumulate) operations. CiM improves the throughput and the power of AI calculations by over 100x. CiM will be heterogeneously integrated with traditional processors such as CPUs. To operate high capacity CiM as a new accelerator and realize a heterogeneous integration with CPUs, software and hardware (CiM device) co-optimization is essential.