Improving efficiency of ReRAM-based accelerators for cognitive computing workloads

Dec 9, 2025 by Mohammad Sabri Abrebekoh

DOI 10.5821/dissertation-2117-450500

We built ReDy, ARAS and Hamun to make ReRAM-powered PuM accelerators actually practical: ReDy cuts ADC energy by dynamically bit-serial quantizing activation groups on-the-fly, ARAS lets huge models run on tiny ReRAM budgets by smart scheduling, overlapped writes and weight re-encoding, and Hamun stretches device lifetime with fault-aware retirement, wear-leveling and batch execution — together they save energy, boost throughput and push ReRAM inference from toy demos toward real deployments.

source S2, crossref, openalex