Binh ports PhysX to AMD Radeon and wins the global AMD robotics hackathon

TL;DR

Binh (LiveKit) ported NVIDIA’s PhysX 5 GPU solver — the CUDA-only rigid-body core underneath ManiSkill, Isaac Gym, and Isaac Lab — to run on AMD Radeon GPUs via ROCm, and used the port to win the global AMD robotics hackathon (Track 3). The winning demo re-runs the SO-Frame Sim2Real manipulation pipeline entirely on an AMD W7900, replacing the RTX5090 baseline. Reported ManiSkill Franka PushCube throughput is 67,320 steps/s on the W7900 versus 110,705 steps/s on the RTX5090 at 4096 parallel envs — a 1.67× gap that Binh attributes primarily to the 5090’s 2× memory bandwidth, since the sim step is memory-bound. Full Sim2Real training curve reproduces on the W7900 and the trained policy transfers to the real rig.

Key claims

Most published manipulation benchmarks and tuned task suites (ManiSkill, Isaac Gym, Isaac Lab) sit on the CUDA-only PhysX GPU solver and had no AMD path prior to this work.
NVIDIA’s release of PhysX 5’s GPU source made a CUDA→HIP port tractable enough to complete inside a hackathon timeframe.
ManiSkill Franka PushCube on W7900 hits 67,320 steps/s at 4,096 parallel envs vs 110,705 on RTX5090 — 1.67× slower, consistent with the 5090’s ~2× memory bandwidth given the memory-bound step.
The full SO-Frame Sim2Real pipeline (rigid-body sim + visuomotor policy training) reproduces end-to-end on the W7900 with a similar training curve to the 5090 baseline, and the policy transfers to the real rig.

Method

The tweet thread describes porting the GPU code path of NVIDIA PhysX 5 (the rigid-body solver that ManiSkill 3 depends on) from CUDA to run on AMD Radeon (presumably via ROCm/HIP). The tweet does not disclose the specific porting strategy (hipify vs manual rewrite) or which subset of PhysX features were mapped. Validation was done by running ManiSkill 3 tasks and comparing steps/s and Sim2Real transfer against an RTX5090 baseline. The submitted deliverable is a hackathon PR to AMD-DEV-CONTEST/Radeon-hackathon-2026-07 (Track 3, robotics simulation).

Results

ManiSkill 3 Franka PushCube, 4096 parallel envs: 67,320 steps/s on Radeon W7900 vs 110,705 steps/s on RTX5090 (1.67× ratio).
Sim2Real transfer for SO-Frame reproduces on W7900 with a training curve comparable to the RTX5090 baseline; policy successfully drives the real rig (video in tweet).
No third-party replication or throughput comparison against Genesis / MuJoCo on AMD is provided.

Why it’s interesting

Every filed RL-environment platform in RL Environment Platforms that runs on GPU physics (ManiSkill, Isaac Lab, RoboLab, Anyscale Ray + Isaac Lab, RoboCasa365, ABC Sim) is CUDA-bound today; the platform’s TL;DR explicitly names the CUDA-only PhysX solver as the load-bearing substrate. This tweet is the first filed datapoint suggesting that lock-in isn’t permanent — a single hackathon submission put a working AMD path under the same benchmarks. Complements Genesis World — Simulation platform for general-purpose robotics & embodied AI learning (Genesis World was designed AMD-friendly from the start) and The Next Leap: AMD MI455X (AMD MI400 hardware supply) by supplying the missing software-porting datapoint. If it replicates, the practical implication for the concept is that env-compute pools no longer need to be homogenous NVIDIA fleets, which changes the cost model for the “decouple env compute from training compute” design pattern already established there.