A big letdown for me is, except with some rare cases, those extra AI features useless outside of AI. Some NPUs are straight out DSPs, they could easily run OpenCL code, others are either designed to not be able to handle any normal floating point numbers but only ones designed for machine learning, or CPU extensions that are just even bigger vector multipliers for select datatypes (AMX).
A big letdown for me is, except with some rare cases, those extra AI features useless outside of AI. Some NPUs are straight out DSPs, they could easily run OpenCL code, others are either designed to not be able to handle any normal floating point numbers but only ones designed for machine learning, or CPU extensions that are just even bigger vector multipliers for select datatypes (AMX).