The energy cost associated to moving data across the memory hierarchy has become a limiting factor in modern computing systems. To mitigate this trend, novel computing architectures favoring a more local and parallel processing of the stored information are proposed, under the labels « Near/In-Memory Computing » or « Processing In Memory ». Substantial benefits are expected in particular for computationally complex (e.g. combinatorial optimization, graph analysis, cryptography) and data-intensive tasks (e.g. video stream analysis, bio-informatics). Such applications are especially demanding in terms of endurance, latency and density. SRAM, fulfilling the first two criteria, may eventually suffer from its footprint and static power consumption. This prompts the evaluation of alternative denser and non-volatile memory technologies, with magnetoresistive memories (MRAM) currently leading in terms of speed-endurance trade-off.
The primary objective will be to estimate improvements brought by MRAM in terms of array-level power, performance, area (PPA), as compared to SRAM-based on-chip memories and for advanced technology nodes. The candidate will establish an analysis and benchmarking workflow for various classes of MRAM, and optimize single bit cells based on a compact model for the memory element. This baseline approach will then be adapted to functional variations specific to IMC in order to assess the benefits of MRAM on an integrated test vehicle.