In a previous post, I used Intel’s AMX instructions intended for AI/ML use-cases to take the average color of an image. This was primarily a proof-of-concept since pedestrians like me generally don’t have access to Intel’s’ AMX-enabled hardware. The cost-of-entry for Intel’s Sapphire Rapids chips is pretty steep too. Maybe some day it will be ubiquitous in consumer-hardware and share a similar story as AVX-512.
Pedestrians like me do have access to an Apple M2 Mini though, after some frustration with trying to sustain a development-environment with a MacOS VM:
These chips, confusingly, also have an instruction set referred to as “AMX” for AI/ML use-cases, sharing the same name-space as Intel’s AMX instructions.
This instruction set has been reverse-engineered as an open-source effort from the Apple-Clang compiler symbols, the iOS SDK, and other sources.