Sporkle: putting memory and device orchestration first
An uncomfortable truth: our CUDA based AI stacks were built by people who optimize equations, not hardware.
Lynn Cole / 标签
A thread through the work.
An uncomfortable truth: our CUDA based AI stacks were built by people who optimize equations, not hardware.
The first GPU implementation worked. Then a CPU benchmark exposed a regression: a novel memory-management system had been replaced by seven nested loops.