Nobody recompiles PTX in compute: cubin scheduling is baked into the binary
mike64_t · x · 2026-11-01
mike64t lays out the CUDA compilation model: fat binaries ship machine code for every targeted sm version with no final recompilation; 'virtual sm' is just a hint that an sm100 binary originated from sm90 PTX. In practice nobody in compute retargets PTX — cubins are the actual SM machine-code ELF binaries the GPU runs, with instruction-level scheduling baked in. QMD/nvpushbuf volatility is a separate matter, unrelated to this fingerprintable GPU nondeterminism.
Related event: GPU Experts Debate Whether Driver Versions Break CUDA Atomic Determinism(9 posts)→
More from Infra
- Local MiniMax H3 video gen on RTX 5060 8GB: 27 tests charted, gen time scales steeply with resolution — rguerraf · 2026-11-01
- China's Halloween cosplay peak: someone dressed up as Intel's -28.58% stock crash chart — basedjensen · 2026-11-01
- Is llama.cpp the new Ollama? Reddit debates the fragmented local LLM engine scene — fintip · 2026-11-01
- Researcher teases talk synthesizing a decade of large-scale training systems — _arohan_ · 2026-11-01
- vLLM cold starts on RunPod cut to 14s: DMTCP snapshots where CRIU fails — kbesplv · 2026-11-01
- Meme: writing an LLVM backend for a fully undocumented GPU architecture to ditch closed-source nvcc — mike64_t · 2026-11-01