AI Reveals New Possibilities in Matrix Multiplication
quantamagazine.org
quantamagazine.org
https://bugfix-66.com/2d447332bf8f8c67348d520c6508b2cc7d3204...
The basic idea is that "smart memories" can be used to do make the multiplication very local and parallel.
The above code spreads matrices out within 64-bit registers, but you could generalize this to huge, multi-megabyte registers.