PyTorch nested graph breaks deliver 4.53× steady-state speedup on B200
cfregly · x · 2026-10-26
A developer benchmarked PyTorch 2.14.1 (CUDA 13.0) with nestedgraphbreaks=True on an NVIDIA B200, reporting a 4.53× steady-state speedup, with benchmark scripts, 24 runs, and graphs open-sourced on GitHub.
Key findings:
- 4.53× steady-state speedup with torch.compile
- First 10 compiled-call intervals were 33.2% slower with the flag on (compilation overhead)
- Benchmark code and analyzer available in the ai-performance-engineering repo
Takeaway: best for long steady-state workloads; weigh the cold-start compilation cost.
More from coding & agent
- Resin: a runtime that lets agents write forward passes to build standalone LLM inference engines — andersonbcdefg · 2026-10-26
- Developers share the hard rules they added to agents after a single bad experience — Luvena21 · 2026-10-26
- unified-brain: OSS SQLITE memory layer claims frontier-class coding at 1/18th cost — StatusFoundation5472 · 2026-10-26
- Dev shares multi-agent coding stack: routing tasks to Opus, Fable and Astra by code size — sh_reya · 2026-10-26
- App Store Connect CLI 5.17.0 ships one-call screenshot reordering, 8k stars on GitHub — rudrank · 2026-10-26
- Dev shows GPT-6 Astra Ultra coding for 41+ hours straight on a Dreamcast game — AIandDesign · 2026-10-26