Hi NVIDIA team,
I’m trying to run an Isaac Sim ROS 2 simulation with renderer multi-GPU enabled.
Basic Isaac multi-GPU works in the container, but the simulation consistently
segfaults when RTX lidars are enabled.
The failure appears specific to this combination:
RTX lidar + Isaac renderer multi-GPU
Environment
- Isaac Sim: 5.1
isaacsim.sensors.rtx: 15.8.4omni.sensors.nv.lidar: 3.0.0- Driver: 575.57.08
- Graphics API: Vulkan
- OS: Ubuntu 22.04.5, kernel
6.8.0-110-generic - Running inside Docker/container
- ROS 2 Humble with Isaac ROS 2 bridge
- GPUs: 2x NVIDIA GeForce RTX 5090, 32 GB each
Inside the container CUDA sees both GPUs.
SimulationApp config
The failing run starts Isaac roughly like this:
SimulationApp({
“headless”: True,
“renderer”: “RaytracedLighting”,
“multi_gpu”: True,
“max_gpu_count”: 2,
“fast_shutdown”: True,
“extra_args”: [
“–/renderer/multiGpu/activeGpus=0,1”,
],
})
What works / fails
I tried to isolate this carefully:
| Test | Result |
|---|---|
| Minimal headless SimulationApp with multi_gpu=True, active GPUs 0,1 | Passes |
| Same scene, cameras enabled, lidars disabled, multi-GPU enabled | Passes |
| Same scene, RTX lidars enabled, multi-GPU disabled | Passes |
| Same scene, RTX lidars enabled, multi-GPU enabled | Segfaults |
So the GPUs, Vulkan, and basic Isaac multi-GPU path appear to work. The crash
only appears when RTX lidar is active while renderer multi-GPU is enabled.
Crash stack
The crash is a native SIGSEGV, exit code -11. The relevant backtrace is:
libcuda.so.1!cuLaunchKernel
libomni.sensors.nv.lidar.rtxmodel_lidar_core.plugin.so!
omni::sensors::materials::MaterialManager::processReturns(
omni::sensors::materials::MaterialManager::LaunchCfg const&, void*
)
libomni.sensors.nv.lidar.rtxmodel_lidar_core.plugin.so!
omni::sensors::nv::lidar::processBatchReturn(…)
librtx.rtxsensor.plugin.so
libgpu.foundation.plugin.so
libcarb.tasking.plugin.so
The important frame seems to be:
omni::sensors::materials::MaterialManager::processReturns
This suggests the crash happens during RTX lidar/material return CUDA processing,
before the data is successfully published through ROS.
Workarounds tried
I tried disabling GPU output buffers for lidar:
--/app/sensors/nv/lidar/outputBufferOnGPU=false
That did not change the crash.
I also tried changing the ROS RTX lidar helper fullScan setting. That also did
not change the crash, which makes sense because the segfault appears to happen
inside the NVIDIA RTX lidar plugin/material return path rather than in ROS
publication.
Question
Is RTX lidar expected to work with Isaac renderer multi-GPU in Isaac Sim 5.1?
If yes, is there a required setting for binding RTX lidar processing to a
specific GPU, disabling cross-GPU lidar processing, or forcing the lidar return/
material processing buffers onto the same device as the render product?
If no, it would be useful to know whether this is a known limitation/bug of RTX
sensors with renderer multi-GPU, especially on multi-GPU RTX 50-series systems.