Exploiting Subword Permutations to Maximize CNN Compute Performance and Efficiency

authored by

Michael Beyer, Sven Gesper, Andre Guntoro, Guillermo Paya-Vaya, Holger Blume

Abstract

Neural networks (NNs) are quantized to decrease their computational demands and reduce their memory foot-print. However, specialized hardware is required that supports computations with low bit widths to take advantage of such optimizations. In this work, we propose permutations on subword level that build on top of multi-bit-width multiply-accumulate operations to effectively support low bit width computations of quantized NNs. By applying this technique, we extend the data reuse and further improve compute performance for convolution operations compared to simple vectorization using SIMD (single-instruction-multiple-data). We perform a design space exploration using a cycle accurate simulation with MobileNet and VGG16 on a vector-based processor. The results show a speedup of up to 3.7 × and a reduction of up to 1.9 × for required data transfers. Additionally, the control overhead for orchestrating the computation is decreased by up to 3.9 ×.

Organisation(s)

Institute of Microelectronic Systems

External Organisation(s)

Robert Bosch GmbH
Technische Universität Braunschweig

Type

Conference contribution

Pages

61-68

No. of pages

8

Publication date

2023

Publication status

Published

Peer reviewed

Yes

ASJC Scopus subject areas

Hardware and Architecture, Computer Networks and Communications

Electronic version(s)

https://doi.org/10.1109/ASAP57973.2023.00023 (Access: Closed)