VSCNN: Convolution Neural Network Accelerator With Vector Sparsity

TL;DR

This paper introduces VSCNN, a hardware accelerator that efficiently supports both dense and vector sparse CNNs using a unified design, achieving significant speedup over traditional dense CNN accelerators.

Contribution

The paper presents a novel CNN accelerator supporting vector sparsity with low overhead, enabling flexible and efficient processing of both dense and sparse networks.

Findings

Achieves 1.93X speedup over dense CNN accelerators

Supports both dense and vector sparse CNNs with the same hardware

Reduces control complexity compared to fine-grained sparse support

Abstract

Hardware accelerator for convolution neural network (CNNs) enables real time applications of artificial intelligence technology. However, most of the accelerators only support dense CNN computations or suffers complex control to support fine grained sparse networks. To solve above problem, this paper presents an efficient CNN accelerator with 1-D vector broadcasted input to support both dense network as well as vector sparse network with the same hardware and low overhead. The presented design achieves 1.93X speedup over the dense CNN computations.