15、Sparse-PE: Revolutionizing Sparse CNN Acceleration

Sparse-PE: Revolutionizing Sparse CNN Acceleration

1. Introduction

In the realm of Convolutional Neural Network (CNN) inferences, acceleration has been a focal point of research. Various architectures have been proposed to enhance the efficiency of CNN computations, with a particular emphasis on exploiting sparsity to reduce compute and memory access volume. This blog explores the limitations of existing dense and sparse architectures and introduces a novel multithreaded Processing Element (PE) called Sparse - PE, which addresses these limitations and offers improved performance.

2. Related Work

2.1 Dense Architectures

• Compute Optimization : Some accelerators focus on optimizing compute operations. For