PufferLib 5.0 hits 60M steps/sec single-GPU RL training, solves Breakout in under a second
jsuarez · x · 2026-09-17
PufferLib 5.0 ships alongside an in-depth engineering writeup: up to 60M useful training steps per second on a single GPU, enough to solve Breakout in under a second. Version 4.0's custom CUDA C stack took an 8-month slog; 5.0 polishes the jank off.
Key numbers: throughput scales linearly down to below 100k parameters — 50M sps at 100k params, 10M sps at 1M params, and 1.8M sps at 10M params. The author argues this reverse-direction scaling (small-model efficiency, not the communication-overhead-besieged scaling-up race) is what makes the result striking.
More from Infra
- Palmer Luckey asks: can you use datacenter RAM for things prohibited by the operator's terms of service? — avaitopiper · 2026-11-02
- Disable MTP for 20-30% faster local LLM inference on limited RAM — InternalMode8159 · 2026-11-02
- Kioxia up 216% in a year but down 68% from its June 2026 peak — zephyr_z9 · 2026-11-02
- Intel plunges 29%, worst day since the 1970s, wiping $12B off US government stake — emmanuelvivier · 2026-11-02
- Researcher hoards nearly 1TB of DRAM, refuses to sell a stick to data centers — kevinnbass · 2026-11-02
- Self-Hosting LLMs at a Law Firm: Infra and Data Governance Harder Than the Model — EducationalGroup7245 · 2026-11-02