Reducing GPU memory usage for multi-camera uses

[wchen]

I’m currently trying to simulate 8 or so cameras at once while having them all publish with ROS. However, it seems like doing so causes my GPU to run out of memory (RTX 2080, 8 GB). I’ve found that I can get up to around 4 cameras working at the same time, using up ~7.2 GB. Are there any methods for lowering the memory usage for many-camera use cases?

EDIT: I thought I should mention that I am effectively only using the cameras for ROS. If there’s a way to disable them rendering to viewports while still updating in ROS, that might also help.

[ltorabi]

Hi wchen,

We are aware of the memory usage of the multi camera, and also about the limitation that the camera is tied to the viewport. We are working on it for the next release.

Kindly,
Liila

[wchen]

Thanks! Do you guys have an estimate for when the next release is?

[ltorabi]

Next release will be in October

[fzg7919]

I meet the same issue, multi camera leads to core dump, beacause of not enough GPU memory .

here is the isaacsim app log
kit_20220614_110632.log (3.2 MB)

[ltorabi]

for multi camera with high resolution, you do require multi GPU

[fzg7919]

Thanks ! but how to set multi camera with multi GPU ? I run my robot in isaacsim container on remote headless machine with 8 X 3080TI ? It runs normally but the omnigraph’s publisher rate is very low , less than 10hz

[ltorabi]

Hopefully this will help you: Rendering Basics — Omniverse Materials and Rendering documentation

[loic.sacre]

We have 2 GeForce RTX 3080. Having only 6 robots, each with a camera (with which we capture a depth image), results in taking half of the memory of each GPU. The camera resolution is only 171 x 224… Is it normal that it takes so much GPU RAM ? Or are we doing something wrong ?

Without the cameras, we can easily have 100 robots.

Here are some memory snapshots (from nvidia-smi):

6 robots:

11 robots:

21 robots:

26 robots: Out of memory