ISCA 2018 / neural-network acceleration

UCNN: Exploiting Computational Reuse in Deep Neural Networks via Weight Repetition

Kartik Hegde, Jiyong Yu, Rohit Agrawal, Mengjia Yan, Michael Pellauer, and Christopher W. Fletcher

Diagram comparing UCNN computation reuse with a conventional neural-network accelerator — UCNN reuses arithmetic associated with repeated neural-network weights.

Scope

Quantized neural networks repeat many weight values, not only zero. UCNN reuses partial computations associated with those repeated weights.

Abstract

CNN inference consists largely of dot products between weights and activations. UCNN generalizes sparsity-based acceleration by treating every repeated weight as an opportunity for reuse. Algebraic reassociation reduces repeated computation and compresses model storage, while conventional sparsity optimizations remain available. Evaluation across three CNNs shows 1.2×–4× lower throughput-normalized energy with a 17–24% processing-element area overhead.