Developing a BLAS Library for the AMD AI Engine [pdf] (uni.tlaan.nl)

A new research thesis from Tristan Laan at Vrije Universiteit Amsterdam introduces "aieblas," an innovative Basic Linear Algebra Subprograms (BLAS) library designed specifically for AMD's AI Engine (AIE). This development aims to simplify programming for spatial dataflow architectures, which traditionally require complex programming languages and deep device-specific knowledge. By enabling high-level performance through a more generalized library, aieblas allows developers to execute linear algebra operations on the AIE without needing expertise in the underlying hardware or software tools. This advancement is significant for the AI/ML community as it enhances the usability of AI accelerators by providing robust support for general computations, potentially broadening the appeal and accessibility of AIEs for various applications, including high-performance computing (HPC). The thesis also evaluates aieblas against OpenBLAS, a widely-used CPU implementation, highlighting its performance and optimizations like dataflow and kernel generation strategies. This research not only contributes to AMD's software ecosystem but also outlines a framework that could potentially be adapted for other spatial dataflow architectures, thus driving future innovations in AI hardware usage and efficiency.