Work

Reducing Off-chip Memory Accesses in Deep Neural Network Accelerators

Year: 2019
Type: dissertation
Abstract: The popularity of deep neural networks (DNNs) has led to widespread development of specialized hardware for accelerating their computations. Many designs employ a memory system that consists of an on-chip local memory in addition to the off-chip main memory. In this dissertation, I seek to improve the energy efficiency in DNN accelerators by reducing the off-chip memory accesses. In the first part, I present and evaluate the required on-chip memory capacity and off-chip data footprint of various configurations for the on-chip memory. For memory configurations where off-chip memory accesses are reduced to once per value, up to 14.3 MB of on-chip memory capacity is necessary. In the second part, I introduce a compression scheme that can be added to any DNN accelerator. It leverages the inherent bit sparsity in DNNs to compress the weight values. Doing so, the off-chip data footprint is reduced by 2.5x-7.6x across the DNNs studied.
Author: Kevin Siu
Language: English
Open Access status: green
OpenAlex ID: W3160069499
Cites: 47
Cited by:
Related to: 10