Improving memory-centric architectures for accelerating cognitive computing workloads

Dec 10, 2025 by Bahareh Khabbazan

DOI 10.5821/dissertation-2117-450564

We rethought the memory hierarchy for DNNs and built a suite of memory-centric tricks and accelerators — DNA-TEQ for adaptive exponential quantization that slashes footprint and nukes multipliers, QeiHaN for base-2 PnM with implicit bit-shifts that collapses data movement, and Lama/LamaAccel to enable parallel column-independent DRAM LUTs and HBM-backed LLM runs — together they cut memory traffic, speed up inference massively, and make large models far more energy efficient without exotic DRAM mods.

source S2, crossref, openalex