- 
                Notifications
    You must be signed in to change notification settings 
- Fork 2.5k
Description
Bug Description
(this is more of an FYI, feel free to close it if it's within acceptable limits)
I believe the differentiable constraint solver feature (commit 44e0879) increased the memory footprint of genesis by about 14%. For simple environments, this isn't a big problem. However, for one of my slightly more complex robots it reduces the number of parallel simulations by about 1,000.
For reference, though, I'm still able to run far more parallel simulations using Genesis than Isaac Lab. So either way, your team is doing great work!
Steps to Reproduce
Run the Genesis locomotion example and check the average GPU memory usage.
On my system
Before this commit: 3573MiB
After this commit: 4082MiB
Expected Behavior
The memory usage shouldn't go up quite so much. Then again, there are always tradeoffs for better features, like making the rigid body simulation to be differentiable. So maybe this is well within the expected increases as the system matures.
Screenshots/Videos
No response
Relevant log output
Environment
- OS: Ubuntu 24
- GPU/CPU: RTX 5070 Ti
- GPU-driver version: 580.82.07
- CUDA / CUDA-toolkit version: 13.0
Release version or Commit ID
Latest main branch, commit 345aa3c
Additional Context
No response