Maximizing Multi-Core Efficiency in BLAS: A Scalable Architecture for Performance

Basic Linear Algebra Subprograms (BLAS) operations are fundamental computational building blocks widely used in machine learning models and AI libraries. Among BLAS routines, General Matrix multiplication(GEMM) is one of the most computationally intensive tasks, making it crucial to leverage multi-core CPU systems for enhanced efficiency and speed. The existing threading architecture in OpenBLAS, a popular BLAS library, only allows for a single execution of GEMM at a time, which fails to meet the extensive computational requirements of modern applications. This limitation restricts parallelism and CPU core utilization.This poster introduces a scalable and enhanced architecture in OpenBLAS that facilitates parallel BLAS executions with loosely coupled synchronization using POSIX conditional variables, thereby boosting performance and core utilization. With the proposed algorithmic changes to OpenBLAS, we have observed an average improvement of 30 % for matrix sizes of 128 and 256 on a 64-core AWS Graviton3 instance.