heterogeneous compute

A thread through the work.

Sporkle: putting memory and device orchestration first

An uncomfortable truth: our CUDA based AI stacks were built by people who optimize equations, not hardware.