Compute-in-ROM : a in-memory activation-aware dot-product compute macro for low-cost edge devices

Access full-text files

Date

2021-05

Authors

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Compute-in-Memory (CiM) approaches perform multiply-and-accumulate (MAC), or dot-product, arithmetic operations required for neural network (NN) within the memory array that stores weight. Many CiM works within different memory technologies have reported superior energy efficiency compared to conventional digital accelerators. To overcome reliability and cost issues that are present in other memory arrays, this report details a CiM approach within in Read-Only Memory (ROM) that features (1) multi-level cell (MLC) for dense storage, (2) accumulation performed through current summation that has improved robustness to process variation, (3) capacitor- and resistor-less compact peripheral to enable analog compute, and (4) ReLU activation-aware MAC compute for early termination of computing negative pre-activation to achieve better energy efficiency.

Description

LCSH Subject Headings

Citation