Handling multiple PxScene with single PxPhysics

[SumantBagri]

Hello NVIDIA Devs,

Happy Holidays!!

I am currently working on a project where I am planning on creating multiple PxScene instances with a single PxPhysics instance. In this particular case, I would also like to enable the GPU rigid-body-pipeline on all the PxScene instances. I wanted to know the following:

1. If I have N * X rigid bodies, is it better to have X objects distributed across N, GPU-enabled PxScene instances or have N * X objects in a single GPU-enabled PxScene. Better in terms of GPU memory access patterns that would ultimately affect the performance
2. If there is a snippet that manages multiple GPU-enabled PxScene instances, could you point me to it ?
3. If there are any specific details that I should be aware of while maintaining multiple, GPU-enabled PxScene instances ?

Thank you in advance!

Hope you have an amazing holiday season :)

[preist-nvidia]

Hi @SumantBagri,

Happy holidays to you, too!

As so often, the answer is "it depends" on your specific simulation setup, and you should try both approaches and measure memory / perf and compare. But I asked the team and we would think that:

• In most cases you will be faster with a single scene for N*X rigids.
• The multi-scene approach will use more memory than the single-scene approach.

We don't have a multi-GPU-scene snippet, unfortunately. I am not aware of any special details you should be aware of for multi-scene GPU setups.

Please update us about what you find, or if you run into issues!

Cheers,
Philipp

Internal tracking: PX-4670