Random-action physics performance per USD
RaiSim on Intel Core Ultra 7 265K versus Newton on GeForce RTX 5070. Independent ANYmal joint-position targets and Strandbeest crank-speed targets, refreshed every ten physics steps in every environment. One world and robot per environment.
How the price difference affects value
The comparison divides each engine’s throughput by its own component price. A value ratio above 1 favors RaiSim. The ratio is performance per dollar, not the hardware-price ratio or the raw-throughput ratio.
Random-action protocol
- ANYmal: 12 PD position targets uniformly within ±0.15 rad of the nominal pose; Kp 200, Kd 10. Strandbeest: one crank velocity target, ±25% around 2π rad/s; Kd 200, linkage joints passive.
- Seed 20261115, keyed by world, control interval and canonical joint. Targets are identical float32 values on CPU and GPU. New actions every ten physics steps: ANYmal 20 ms, Strandbeest 10 ms.
- Includes random generation, target updates and physics. CPU workers synchronize at each control interval; each worker integrates a world ten times before visiting the next. GPU actions and simulation are stream-ordered in captured graphs. No learned policy, observations, rewards, episode resets or rendering are timed. Divide physics FPS/USD by ten for control transitions/s/USD.
- Nominal warmup: 1,100 ANYmal steps or 100 Strandbeest steps; state, action counters, targets and solver warm starts reset outside timing. Three trials per case, except single-world RaiSim uses seven; median elapsed time, no exclusions. All world counts use 5,000 ANYmal or 3,000 Strandbeest steps.
- Untimed diagnostic replay measures contacts and validation. CPU loop gaps and finite states sampled every ten steps; GPU loop gaps/overflow checked every physics step, finite state at the end. Both use a 1 mm Strandbeest gap threshold. Full trial timings, contacts and quality results are embedded in the dataset.
- All measurements are fresh 15 November 2026. RaiSim physics uses FP64, Newton SolverMuJoCo / MuJoCo Warp uses FP32. Same physical source models and controller gains; differing constraints, precision and contact trajectories. These are workload comparisons, not identical numerical work.
- Ten primary CPU cases through 4,096 worlds; larger GPU batches have no measured CPU counterpart. GPU memory safeguard stops further doubling once sampled process usage exceeds 5,000 MiB; this is not a solver world-count limit.
- Earlier fixed-target measurements used a different CPU driver and trial horizons. Changes include synchronization, action updates and dynamic workloads, so cross-workload differences do not isolate random-action cost.
Source: /tmp/raisim-random-actions-20261115/results.json. Full dataset and trial timings are embedded in this HTML.