PufferLib 5.0 trains a Rubik's Cube solver at 8.8M SPS on a single 12GB GPU
haydendevs · x · 2026-11-11
A developer posted a screenshot of PufferLib 5.0 in action: training a Rubik's Cube environment on a single 12GB card at 8.8 million steps per second, with the GPU pegged at 100% while using only 1.8GB of VRAM.
Key stats from the panel: 280.5M steps, epoch 1070, 621ms per evaluation; perf and score both at 1.000, shortest solve 29 steps, policy and value losses near zero. The caption is just "uhhhhh guys"—clearly stunned by the throughput. PufferLib is the well-known extreme-speed RL training library, and this screenshot is its best advertisement.
More from Infra
- Reuters: Chinese military universities obtained restricted Nvidia AI chips despite US export controls — Polymarket · 2026-11-12
- Arm exec on the Agentic CPU: GPUs think, CPUs act — agents need their own class of silicon — BenBajarin · 2026-11-12
- Meta ships KernelEvolve, following Google's AlphaEvolve AI kernel-optimization play — RichardsonDx · 2026-11-12
- AlphaEvolve breaks 56-year-old 4x4 matrix multiplication record, cutting 49 to 48 — RichardsonDx · 2026-11-12
- Scoop: Inference providers Fal and Baseten move to buy their own chips as clouds prioritize big customers — steph_palazzolo · 2026-11-12
- DeepSeek V4.1 Flash runs at 20 tok/s on a single RTX 3090 with 64GB RAM — casper_hansen_ · 2026-11-12