Triton for MTIA: Bridging the Programming Model Gaps for Custom AI Accelerators (arxiv.org)

Meta has announced the successful application of Triton—a high-level kernel programming language—on its custom machine learning accelerator, MTIA-2i, marking a significant milestone in bridging programming model gaps in AI hardware development. As custom accelerators proliferate to handle the increasing demands of machine learning workloads, the need for a unified programming resource has become critical. Triton, alongside TorchInductor, previously demonstrated prowess with GPUs, but its adaptation to the distinctly different programming models of custom accelerators was still in question. In this groundbreaking implementation, researchers developed a new compiler backend specifically for MTIA-2i, optimizing code generation and introducing minimal language extensions to leverage the architecture's unique features. The results reveal that Triton-MTIA kernels perform competitively with expert-tuned C++ implementations, efficiently serving about 60 different model types—accounting for nearly half of the layers and a significant portion of execution time. This achievement illustrates how domain-specific languages (DSLs) like Triton can facilitate rapid model iteration, paving the way for enhanced innovation and scalable deployment in the AI/ML landscape.