Matrix multiply-accumulate operations, possibly at reduced precision (8- or 16-bit). The most common name for this is NPU but the description applies to other things too (eg NVIDIA’s tensor cores, Google’s TPUs, Apple’s ANE to a lesser extent, etc)