RaiSim vs Newton benchmark: CPU robot sim beats GPU by up to 95x at practical batch sizes
ChongZzZhang · x · 2026-11-15
The RaiSim team published a physics benchmark against the GPU-based Newton engine, showing that massive parallelism doesn't automatically make robot simulation fast.
Key findings:
- At recorded hardware prices, RaiSim delivers 2x or more physics performance per dollar at every measured scale; with 1–64 worlds the advantage reaches 9.89–253.51x on ANYmal and Strandbeest benchmarks
- A single CPU worker (265K, $340) is 95.77x faster than Newton on RTX 5070 ($900) for random-action ANYmal; Newton only wins at GPU-only scales of 8,192–32,768 worlds
- Why: CPU worlds schedule independently under uneven workloads (contacts, resets, solver work), while divergent GPU work and batch synchronization leave resources idle
- For rollout-based learners like PPO, RaiSim's small-batch speed lets you size rollouts for the algorithm and update policies more frequently instead of batching to fill the GPU
More from Embodied
- Researcher teaches a Unitree G1 humanoid to ride a mechanical bull after CoRL — RemiCadene · 2026-11-15
- PrimeSim: world model trained on 30k hours of robot demos runs near real-time on one A100 — jiqizhixin · 2026-11-15
- Viral claim: Claude runs a robot arm on a real factory task with zero training — but the model name doesn't exist — IgorCarron · 2026-11-15
- NEO-VLA ditches visual encoders for robot actions, hitting 90.9% on LIBERO frozen — 机器之心 · 2026-11-15
- Zipline's Keller Cliffton on 3M drone deliveries and what breaks at 10x scale — a16z Podcast · 2026-11-15
- AI Tamagotchi: CMU student's E Ink gadgets turn objects into needy pets — Wired AI · 2026-11-15