Programmers love mind-bogglingly complicated weak memory models

on YouTube ↗19m

" summary

Nvidia documents a concrete, unusually weak GPU memory model that adds proxies and enforces proxy preserved program order so that cycles in pppo and fr are forbidden; that design lets texture loads (different indexing/caches) race with regular stores and produce single-threaded incoherence unless a fence is used. Tooling includes Envy litmus that compiles CUDA to PEX and checks outcomes against the PEX axiomatic model. The model prioritizes performance, backwards compatibility, and formal axiomatic proofs as engineering checks, not philosophical purity.

AI-written summary · about

" chapters (6)

1. 0:00Nvidia's weak memory model
2. 2:19Single-threaded data races and fences
3. 4:50Axiomatic vs operational models
4. 7:40Guardrails and tooling for weak memory
5. 10:49Balancing formalism and practice
6. 13:21Conclusion and future