scratch that at 1.8M SPS now
applying the idiot index to everything works so well
training a g1 at 750k SPS on a RTX 6000 Pro
scratch that at 1.8M SPS now
applying the idiot index to everything works so well
training a g1 at 750k SPS on a RTX 6000 Pro
Users are enthusiastic about single-GPU AI training reaching 1.8M samples per second with the Idiot Index metric because it shows pufferlib enabling high-performance robotics simulations.
Just FYI: the fastest SPS I could find online was ~100k about a week ago with a single GPU
Let's keep on pushing training sim speed so we can accelerate robotics and make the world a more automated place!
scratch that at 1.8M SPS now
applying the idiot index to everything works so well

@JulianSaks Nice! Tell me if you get it to converge. Right now my loop is still ~250k on a 4090. What's your num envs?

@JulianSaks If you have a concrete project to attempt and capital, consider contacting pufferlib for an env developed for you, tell them I sent you

@JulianSaks pufferlib detected

@yacineMTB @JulianSaks Will do

@DanAdvantage Yes!! pufferlib is 🤌🏻

@yacineMTB 32,768. wbu?

@yacineMTB pufferlib sighted