Performance of a Multicore Matrix Multiplication Library

Multicore processors promise dramatic improvements in performance, but their diverse and often unique architectures are a major inhibitor to software adoption. Algorithm libraries that operate at the chip level and are optimized across multiple cores provide the quickest route by which programmers can port or develop highperformance software for multicores. This paper reports on a flexible matrix multiplication library for the Cell Broadband EngineTM (BE) processor that meets or exceeds the performance of known matrix multiplication implementations on the Cell. In addition, the library operates within a larger framework for programming multicores that enables programmers to combine library code with multicore functions they have developed themselves.