Data-intensive workloads of very high dimensionality, such as in machine learning, are both memory-limited and compute-limited, from an architectural perspective. In-memory computing (IMC) is a technique which is being explored to address the memory barrier while also making compute operations more efficient. While IMC has been tried in SRAM memories, embedded non-volatile memories (eNVM) promise high density, low idle-state leakage compared to SRAM, without having to re-write on power-on. This FPO introduces four IMC implementations in magneto-resistive RAM (MRAM) and resistive RAM (ReRAM), providing a sequence for discussion of the challenges in implementing IMC in eNVM.

Adviser: Naveen Verma