CartPole

CartPole#

Viewer Performance Impact Benchmark

Abstract#

This benchmark evaluates the performance impact of Brax Viewer on Brax PPO training using the CartPole environment. We measure training throughput under three scenarios: pure training, viewer backend with rendering disabled, and full real-time visualization.

Methods#

Training Settings#

  • Environment: Custom CartPole with mjx backend

  • Training steps: 2,000

  • Episode length: 300

  • Number of parallel environments: 8

  • Learning rate: 3e-4

  • Entropy cost: 1e-2

Implementation#

The benchmark is implemented in benchmarks/brax/cartpole/time_test_small.py. Running the script prints timing results for three scenarios to the console.

Results#

Hardware

No Viewer

Viewer OFF

Viewer ON

Apple M1 Max (32-core GPU)

20.63s

18.17s

51.54s

NVIDIA RTX 5090

51.81s

33.53s

63.57s

NVIDIA RTX A5000

61.21s

47.11s

106.81s

Note

The “No Viewer” results are slower than “Viewer OFF” because the benchmark script runs “No Viewer” first, and the initial run takes longer may due to JIT (just-in-time) compilation.

Contents