RaiSim delivers more physics performance per dollar
At the recorded hardware prices, RaiSim delivers about 2× or more physics performance per dollar at every scale measured for both engines. With 1–64 worlds, the advantage grows to 11.75–192.43× across ANYmal and Strandbeest: over 10× in both benchmarks.
The gap is largest at smaller, practical batch sizes, where fast simulation supports frequent policy updates without waiting to fill a massive batch.
Why CPU simulation fits the learning loop
- Fast from one environment. One CPU worker is 72.69× faster for ANYmal and 13.48× faster for Strandbeest. At 64–256 environments, RaiSim retains 1.60–15.43× higher throughput.
- Update frequently. Filling a larger rollout batch can delay the next policy update. RaiSim's small-batch speed lets you choose rollout size for learning, rather than GPU occupancy.
- Schedule uneven work. Independent CPU worlds can be scheduled separately as robots, contacts, and resets vary in cost. Divergent GPU work and synchronization can leave resources waiting.
These are physics-throughput measurements. Time to target reward also depends on rollout scheduling, data transfers, inference, and optimization. Read the comparison and sources.
Larger GPU batches
Fresh Newton scaling measurements extend beyond 4,096 worlds using the original ten-step graph and validation protocol. More worlds can improve total throughput until the GPU saturates; after that they increase memory use and lower throughput per world. Missing CPU results are not extrapolated.
See the documentation tables and full scaling JSON for sampled GPU memory and validation results. A conservative memory safeguard stops doubling if process GPU usage exceeds 5,000 MiB; this is not a solver world-count limit.
Measurement details
- Only ten-step RaiSim scheduling is shown: each worker integrates a world ten times before visiting its next world. Newton uses GPU graphs of ten steps. These are simulation steps, not rendered frames.
- CPU workers: 1 for one world, 16 for 64 worlds, 20 for 256–4,096 worlds. Worlds are independent, worker-owned and CPU-pinned; the integration inside each world is single-threaded. These worker counts are recorded configurations, not proven optimal.
- Current RaiSim CPU results: 15 November 2026, FP64. Newton GPU: FP32 SolverMuJoCo / MuJoCo Warp. Results through 1,024 worlds are historical 14 November runs; 4,096 and above were rerun on 15 November for the larger-world sweep. Time steps: ANYmal 2 ms, Strandbeest 1 ms. Setup, warmup, validation and external policy computation are outside timing.
- CPU results use median elapsed wall time across repeated trials. Single-world cases: 3 processes × 7 trials. Multiworld cases: 2 processes × 5 trials, except ANYmal 4,096/20 discards the first two trials per process under the recorded protocol, scoring 6 of 10 raw trials. Processes ran serially. Timed horizons vary by world count.
- Physical scenes/controllers were matched, but solvers, precision, constraint formulations and trajectories differ. ANYmal has four contacts/world. Strandbeest historical trajectory means differ: RaiSim approximately 0.959 contacts/world, Newton 0.686–0.700; current RaiSim ending count is one/world. RaiSim merges fixed links (74 bodies); Newton retains them (92). These value ratios do not measure identical numerical work.
- All 18 measured CPU configurations matched one-step final state hashes and contacts; six extra checks covered partial blocks and uneven worker ranges. This report retains ten primary CPU cases and adds larger GPU batches. No CPU measurement is available above 4,096 worlds.
Source throughput: /tmp/raisim-world-heap10/fps-and-heap.json. Historical GPU source: /tmp/raisim-newton-20261114/comparison.json. Fresh scaling source: /tmp/newton-large-worlds-20261115/results.json. No new physics measurements or library changes were made for this report.