They're designed to run ML specific functions like matrix multiply and stuff. Nvidia has a similar idea in "tensor cores". I think because they're low but opera

They're designed to run ML specific functions like matrix multiply and stuff. Nvidia has a similar idea in "tensor cores". I think because they're low but operations like 8 or 16 bit which is faster but too low res for GPU work.