GPU compute is deterministic — the nondeterminism comes from kernel path selection, says dev
felipesuch · x · 2026-10-30
Pushing back on the "GPUs are nondeterministic" take, felipesuch clarifies: compute is deterministic when inputs are. What users perceive as nondeterminism usually comes from which kernel/execution path the code takes — driven by variables outside your control like batch size and SKU.
Related event: GPU Compute Is Deterministic; the Real Culprit Is Kernel Path Selection(2 posts)→
More from Infra
- AgentX benchmark: lossless-CSF hits 11.4M KV cache tokens, 8.7x DeepSeek V4.1 Flash baseline — TheZachMueller · 2026-10-30
- Dev blasts RunPod: on-demand B200 compute 'unreliable at best' for consistent workloads — Kyrannio · 2026-10-30
- 'Girls only want one thing': an $496,639 8x NVIDIA B300 deep learning server — Kyrannio · 2026-10-30
- SAM3 now runs offline on your GPU: ONNX export hits 0.37s on M5, 0.08s on RTX 3080 Ti — CSProfKGD · 2026-10-30
- Same model, different hardware: dev finds A100s vs B300s produce different outputs — Kyrannio · 2026-10-30
- Small devs can't reliably rent compute: grabbing 8 B300s on Modal is 'a shot in the dark' — Kyrannio · 2026-10-30