Yaroslav Bulatov launches A100-targeted architecture experiments attacking backprop's memory wall

yaroslavvb · x · 2026-09-12

Yaroslav Bulatov argues there's room for AI architectures customized to existing hardware and is experimenting on A100s. His open-source sutro-problems repo poses an MNIST challenge: hit target accuracy while addressing the memory wall — e.g., an A100 kernel measured in few Joules via NVML, or an algorithm with a small footprint on Bill Dally's 2D grid. Motivation: backprop dates from the arithmetic-bound 1980s; the bottleneck today is memory movement, and batching only partly mitigates it.

Original post →

More from Infra

Infra channel →