Performance Benchmarks

Isaac Lab leverages end-to-end GPU training for reinforcement learning workflows, allowing for fast parallel training across thousands of environments. In this section, we provide runtime performance benchmark results for reinforcement learning training of various example environments on different GPU setups. Multi-GPU and multi-node training performance results are also outlined.

Benchmark Results

All benchmarking results were performed with the RL Games library with --headless flag on Ubuntu 22.04. Isaac-Velocity-Rough-G1-v0 environment benchmarks were performed with the RSL RL library.

Memory Consumption

Environment Name		# of Environments	RAM (GB)	VRAM (GB)
Isaac-Cartpole-Direct-v0		4096	3.7	3.3
Isaac-Cartpole-RGB-Camera-Direct-v0		1024	7.5	16.7
Isaac-Velocity-Rough-G1-v0		4096	6.5	6.1
Isaac-Repose-Cube-Shadow-Direct-v0		8192	6.7	6.4

Single GPU - RTX 4090

CPU: AMD Ryzen 9 7950X 16-Core Processor

Environment Name	# of Environments	Environment Step FPS	Environment Step and Inference FPS	Environment Step, Inference, and Train FPS
Isaac-Cartpole-Direct-v0	4096	1100000	910000	510000
Isaac-Cartpole-RGB-Camera-Direct-v0	1024	50000	45000	32000
Isaac-Velocity-Rough-G1-v0	4096	94000	88000	82000
Isaac-Repose-Cube-Shadow-Direct-v0	8192	200000	190000	170000

Single GPU - L40

CPU: Intel(R) Xeon(R) Platinum 8362 CPU @ 2.80GHz

Environment Name	# of Environments	Environment Step FPS	Environment Step and Inference FPS	Environment Step, Inference, and Train FPS
Isaac-Cartpole-Direct-v0	4096	620000	490000	260000
Isaac-Cartpole-RGB-Camera-Direct-v0	1024	30000	28000	21000
Isaac-Velocity-Rough-G1-v0	4096	72000	64000	62000
Isaac-Repose-Cube-Shadow-Direct-v0	8192	170000	140000	120000

Single-Node, 4 x L40 GPUs

CPU: Intel(R) Xeon(R) Platinum 8362 CPU @ 2.80GHz

Environment Name	# of Environments	Environment Step FPS	Environment Step and Inference FPS	Environment Step, Inference, and Train FPS
Isaac-Cartpole-Direct-v0	4096	2700000	2100000	950000
Isaac-Cartpole-RGB-Camera-Direct-v0	1024	130000	120000	90000
Isaac-Velocity-Rough-G1-v0	4096	290000	270000	250000
Isaac-Repose-Cube-Shadow-Direct-v0	8192	440000	420000	390000

4 Nodes, 4 x L40 GPUs per node

CPU: Intel(R) Xeon(R) Platinum 8362 CPU @ 2.80GHz

Environment Name	# of Environments	Environment Step FPS	Environment Step and Inference FPS	Environment Step, Inference, and Train FPS
Isaac-Cartpole-Direct-v0	4096	10200000	8200000	3500000
Isaac-Cartpole-RGB-Camera-Direct-v0	1024	530000	490000	260000
Isaac-Velocity-Rough-G1-v0	4096	1200000	1100000	960000
Isaac-Repose-Cube-Shadow-Direct-v0	8192	2400000	2300000	1800000

Benchmark Scripts

For ease of reproducibility, we provide benchmarking scripts available at source/standalone/benchmarks. This folder contains individual benchmark scripts that resemble the train.py script for RL-Games and RSL RL. In addition, we also provide a benchmarking script that runs only the environment implementation without any reinforcement learning library.

Example scripts can be run similarly to training scripts:

# benchmark with RSL RL
python source/standalone/benchmarks/benchmark_rsl_rl.py --task=Isaac-Cartpole-v0 --headless

# benchmark with RL Games
python source/standalone/benchmarks/benchmark_rlgames.py --task=Isaac-Cartpole-v0 --headless

# benchmark without RL libraries
python source/standalone/benchmarks/benchmark_non_rl.py --task=Isaac-Cartpole-v0 --headless

Each script will generate a set of KPI files at the end of the run, which includes data on the startup times, runtime statistics, such as the time taken for each simulation or rendering step, as well as overall environment FPS for stepping the environment, performing inference during rollout, as well as training.