“MatX's chip architecture features a large systolic array that can be split into smaller pieces to efficiently process workloads like the attention mechanism in transformers, which do not map well to a single large array.”