Dear researchers:
Did you know that your H100 can actually ray-trace like a gaming card? 🎮
for a long time, people have assumed that compute cards like the H100 can only compute, and that you need a graphics card like the RTX 5090 to render. we show that's wrong.
🚨Meet RTX H100.
Blender Cycles on the Open Data benchmark (Blender 5.3.0): ours 48,641. the RTX 5090 is at 19,202, and a stock H100 gets only 7,041. our original goal was to match the 5090. we ended up at 2.5x. our neural rendering approach is lossy. we test within a reasonable loss, and every result passes an image check.
we think nobody doing this before was an accident: graphics people rarely touch compute cards, and people who have compute cards rarely know graphics.
robotic simulators need fast rendering. now a general-purpose compute card can do it, so you don't need heterogeneous compute or a complicated cluster.
we'll extend this to more hardware next. almost ready for GTA 6 on the cluster 🤣
Best,
Wenhao
@sheilsarda The bottle doesn’t need that much: in our tests ~80 N per finger already held this 0.5 kg bottle steady. As shipped, RoboDojo’s control code caps the squeeze near 25 N, and the bottle tips over. If you know a measured X5 grip force, we’d gladly use it.
What does RoboDojo really simulate?
We followed it from the real world to its score, one link at a time. At almost every step, it simulates something other than what it says.
Same bottle, same grasp. As shipped, it tips 78° in the hand. After our fix, it stays level.
@sheilsarda Fair question. We didn’t pick 160 N for the bottle: it’s where the X5’s controller saturates (its open-source SDK’s control law and current trip, via an estimated transmission, ±20–30%, not measured on hardware). A position-controlled gripper told to close squeezes up to that.
@MarioChan2002 Thanks, Tianxing! That makes sense: we’ve seen ourselves how hard interpenetration and PhysX’s other limitations are to tame at benchmark scale. Happy to share our fixes and test cases if they’re useful. Looking forward to the Verified release!
@callumshar@wenhaocha1@andylizf The friction bug affects almost all objects, so technically every task is affected. We haven’t tested yet how model scores will change after we fix every task.
4/
Values no real thing has: five of the arm’s six joints may turn 573° each way. The real arm stops within half a turn. About half the benchmark’s released demonstrations bend a joint further than the real arm can. The garbage is already in the dataset.