Skip to content

Instantly share code, notes, and snippets.

@brycelelbach
Created April 28, 2026 04:32
Show Gist options
  • Select an option

  • Save brycelelbach/22c197670b1ddd3816a15bdc95a30574 to your computer and use it in GitHub Desktop.

Select an option

Save brycelelbach/22c197670b1ddd3816a15bdc95a30574 to your computer and use it in GitHub Desktop.
craftax.cu autocuda results
JAX (upstream, RTX 3090) craftax.cu baseline (upstream, RTX 3090) craftax.c (upstream, Ryzen 9 9950X3D) craftax.c (this box, EPYC 9335) craftax.cu baseline (this box, RTX Pro 6000) craftax.cu optimized (this box, RTX Pro 6000)
Env-only SPS (NE=4096) ~300,000 (1×) 1,400,000 (4.7×) 14,712,099 (49.0×) 1,731,059 (5.8×) 54,177,909 (180.6×)
Env-only SPS (NE=32k) OOM 7,800,000 (1׆) 47,800,000 (6.13׆) 23,103,371 (2.96׆) 12,529,478 (1.61׆) 229,289,626 (29.4׆)
Training SPS (matched hparams) 107,000 (1×) 796,000 (7.4×) n/a n/a 1,148,923 (10.7×) 3,007,291 (28.1×)
Wall time for 10M steps 93.0 s (1×) 12.5 s (7.4×) n/a n/a 8.67 s (10.7×) 3.31 s (28.1×)
Final mean return 4.76 5.48 (same)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment