Robotics StackExchange | Archived questions

choosing a GPU card for a high-end simulation machine

Hi,
I'm considering 4 options for GPU configurations targeted for a strong machine to run gazebo simulations. These are the cards I consider:
1. Nvidia Quadro K6000
2. Nvidia Quadro K5000
3. Nvidia Tesla K20c
4. The K20c + K5000 (maximus configuration)

Does somebody have experience or recommendations about these cards?
What would be the best choice (even if not in this list)?

Thanks!
Eliya

Asked by Eliya on 2014-01-27 09:12:06 UTC

Comments

I don't think Gazebo needs such high end graphic cards, just make sure it's NVidia, and a GPU at around 200$ should be fine.

Asked by AndreiHaidu on 2014-01-27 10:12:26 UTC

the terrain and robot models I run makes the simulation run very slowly (real-time factor ~0.01 with Nvidia quadro 2000). you're right that the rendering is not the major factor here, but computing performance probably is.

Asked by Eliya on 2014-01-28 09:07:09 UTC

Getting a $3k GPU won't resolve that problem though as physics & collision detection runs on the CPU (until Bullet's GPU support is added). If you want to make simulation run faster your best bet is to work on simplifying your collision models and terrain (feel free showing a screenshot of your current model + environment if you can). If you have any laser scanners on your robot you can also test using the GPU variant of the sensor instead of CPU if you're not already doing so.

Asked by ThomasK on 2014-01-28 13:02:28 UTC

Answers

I believe the answers are in the comments. To summarize, a mid-range nvidia card is sufficient since Gazebo is mostly CPU bound.

Asked by nkoenig on 2014-01-29 00:05:22 UTC

Comments