Industry Expert Blogs

A novel AI-acceleration paper presents a method to optimize sparse matrix multiplication for machine learning models, particularly focusing on structured sparsity. Structured sparsity involves a predefined pattern of zero values in the matrix, unlike unstructured sparsity where zeros can occur anywhere. The research was conducted by Democritus University of Thrace (DUTH) in Greece and was sponsored by Codasip University Program.

Structured sparsity has emerged as a promising approach to streamline the complexity of modern Machine Learning (ML) applications and facilitate the handling of sparse data in hardware. Accelerating ML models, whether for training or inference, heavily relies on efficient execution of equivalent matrix multiplications, which are often performed on vector processors or custom matrix engines.

Integrating structured sparsity into existing vector execution

The aim of this study was to integrate the simplicity of structured sparsity into existing vector execution flow and vector processing units (VPUs), thus expediting the corresponding matrix multiplications with minimal redesign in mind. To achieve this goal, a novel vector index-multiply-accumulate instruction is introduced. This instruction facilitates low-cost indirect reads from the vector register file, thereby reducing unnecessary memory traffic and enhancing data locality.

Click here to read more ...