PufferLib 5.0 Hits 60M Steps/Sec on a Single RTX 5090: How 5 Degrees of Parallelism Work
jsuarez · x · 2026-09-14
jsuarez explains how PufferLib 5.0 broke 60M steps/second of RL training on a single RTX 5090 — months of simulated data per second for most environments.
- Environment counts aren't benchmark inflation: hundreds to thousands of envs per GPU, chosen via a 1000+ experiment hyperparameter sweep balancing task performance vs wall-clock time, with results public in the Constellation dashboard.
- Two major parallelization options: thread-based CPU parallelism and more (post truncated), organized around five degrees of parallelism.
He notes they could exceed 100M SPS with a million environments but report maximum throughput under useful configurations.
More from Infra
- Your AI agent would cost 260% more without prompt caching, Arena data shows — arena · 2026-10-30
- Nebius runs auction-driven GPU pricing, lifting rates ~20% amid tight supply to 2028 — BenBajarin · 2026-10-30
- macOS update silently re-adds eMudhra root certs users had removed — steipete · 2026-10-30
- Bittensor's Terza Inference routes one API across miners' models on subnet 103 — const_reborn · 2026-10-30
- SGLang Summit 2026 sells out sponsor booths and Full Passes ahead of Nov 12–13 event — ying11231 · 2026-10-30
- OpenRouter rebuilds provider routing: 46% fewer capacity errors, 16.9% faster TTFT — steipete · 2026-10-30