Simulation Throughput (Steps / Frames Per Second)

仿真吞吐量Advanced

How many environment steps or frames a simulator produces per second, which sets how fast data can be generated and training can run.

Simulation throughput refers to how much interaction data a simulator produces per unit time, usually expressed as steps per second or frames per second (FPS); in GPU-parallel simulation, it is typically reported as the summed steps across thousands of parallel environments. Reinforcement learning often needs hundreds of millions of steps of interaction, so throughput directly determines whether a training run takes days or minutes. NVIDIA's 2021 Isaac Gym put both physics simulation and the neural network on the GPU with no data crossing to the CPU, reporting speedups of 2 to 3 orders of magnitude over the older “CPU simulation plus GPU training” approach. The main factors affecting throughput are the number of parallel environments, the complexity of the model and its contacts, the simulation timestep and number of substeps, solver iteration count, and whether camera images are rendered (rendering is usually much slower). It differs from real-time factor, which measures how much faster a single environment runs than real time. When comparing numbers across papers, check whether they use the same task, the same hardware, and rendering or not.

ExampleETH's legged_gym simulates thousands of ANYmal quadrupeds on a single GPU at once, finishing a flat-ground walking policy in under 4 minutes and a rough-terrain one in about 20; ManiSkill3 reports over 30,000 frames per second with rendering enabled on its benchmark environments.

Also called: Simulation FPS, Steps Per Second
Related: GPU-Accelerated Parallel Simulation · Vectorized Environments · Real-Time Factor · Massively Parallel Reinforcement Learning · Batched Rendering · Sample Efficiency
Sources: arXiv 2108.10470 - Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning
arXiv 2109.11978 - Learning to Walk in Minutes Using Massively Parallel Deep RL
arXiv 2410.00425 - ManiSkill3