Operating System:
Windows
Linux
Kit Version:
110 (Kit App Template)
109 (Kit App Template)
108 (Kit App Template)
107 (Kit App Template)
106 (Kit App Template)
105 (Launcher)
Kit Template:
USD Composer
USD Explorer
USD Viewer
Custom
GPU Hardware:
A series (Blackwell)
A series (ADA)
A series
50 series
40 series
30 series
GPU Driver:
Latest
Recommended (573.xx)
Other 595.91.07
Work Flow:
Main Issue:
Hello,
I am testing Isaac Sim 6.0.0 on an NVIDIA Brev VM with four L40S GPUs. Isaac Sim detects all four GPUs and allocates memory on all of them, but RTX Path Tracing Movie Capture performs rendering work only on GPU 0.
Environment
- Isaac Sim: 6.0.0
- Platform: NVIDIA Brev
- Cloud provider: AWS
- GPUs: 4× NVIDIA L40S
- GPU memory: 46,068 MiB per GPU
- CPU/RAM: 48 vCPUs, 384 GiB RAM
- Driver: 595.91.07
- CUDA: 13.2
- Graphics API: Vulkan
- OS/container: Ubuntu 24.04-based Docker container
- Display: X11 through Brev noVNC
- ROS 2 environment disabled during testing
The VM was created from the Isaac Sim 6.0.0 with ROS 2 Jazzy Extended launchable, but ROS 2 extensions were disabled during launch.
The test scene is:
/workspace/S/1234 GenAI Sunday 12.usd
The same scene opens and renders successfully on a separate single-GPU NVIDIA L40 workstation.
Launch command
/root/isaacsim/isaac-sim.sh \
--no-ros-env \
--/renderer/multiGpu/enabled=true \
--/renderer/multiGpu/autoEnable=true \
--/renderer/multiGpu/activeGpus=0,1,2,3 \
--/renderer/multiGpu/maxGpuCount=4 \
--/rtx/pathtracing/mgpu/enabled=true \
--/rtx/pathtracing/mgpu/autoLoadBalancing/enabled=true
I also tested with:
--/rtx/pathtracing/spp=8
--/rtx/pathtracing/totalSpp=300
GPU detection in Isaac Sim
The Kit log contains this GPU table:
Graphics API: Vulkan
GPU | Name | Active | GPU Memory
0 | NVIDIA L40S | Yes: 0 | 46068 MB
1 | NVIDIA L40S | Yes: 1 | 46068 MB
2 | NVIDIA L40S | Yes: 2 | 46068 MB
3 | NVIDIA L40S | Yes: 3 | 46068 MB
The Isaac Sim viewport HUD also shows memory allocated on all four GPUs, for example:
NVIDIA L40S: 13.3 GiB used
NVIDIA L40S: 12.9 GiB used
NVIDIA L40S: 11.6 GiB used
NVIDIA L40S: 11.6 GiB used
The Render Settings panel shows:
- Multi-GPU: enabled
- Auto Load Balancing: enabled
- Compress Albedo: enabled
- Compress Normals: enabled
- Multi-Threading: enabled
- Compress Radiance: disabled
Movie Capture settings
- Resolution:
3000 × 1687 - Render preset:
RTX - Interactive (Path Tracing) (Current) - Samples per pixel per iteration: tested with
8and16 - Path Trace samples per pixel:
300 - Motion blur: disabled
- Settle latency:
0 - Capture Application: disabled
- Use render product to capture: disabled
- Multi-GPU options enabled
I understand that “Samples per pixel per iteration” is a batching/iteration setting, while 300 is the total path-tracing sample target. I tested 16 samples per iteration with 300 total samples to see whether it would improve GPU distribution.
Actual GPU utilization
During Movie Capture I ran:
nvidia-smi dmon -s pucm -d 1
The output repeatedly showed:
# gpu pwr gtemp sm mem
# Idx W C % %
0 225 59 100 43
1 66 35 0 0
2 65 35 0 0
3 66 35 0 0
This remained consistent throughout capture:
- GPU 0: approximately 100% SM utilization and 225 W
- GPU 1: approximately 0% SM utilization and 66 W
- GPU 2: approximately 0% SM utilization and 65 W
- GPU 3: approximately 0% SM utilization and 66 W
Therefore, GPUs 1–3 appear to have scene memory allocated but are not performing path-tracing work.
Additional log findings
The log registers the Multi-GPU plugin:
Registered plugin: rtx.multigpumanager.plugin
It also maps these settings:
/rtx/pathtracing/mgpu/enabled
/rtx/pathtracing/mgpu/autoLoadBalancing/enabled
/rtx/pathtracing/mgpu/autoLoadBalancingWeightGPU0
/rtx/pathtracing/mgpu/weightGpu0
/rtx/pathtracing/mgpu/printLog
/rtx/pathtracing/mgpu/printWeights
/rtx/pathtracing/mgpu/showCompositingPattern
However, later in the same log it reports:
No setting was found for '/rtx/pathtracing/mgpu/enabled'
No setting was found for '/rtx/pathtracing/mgpu/autoLoadBalancing/enabled'
No setting was found for '/rtx/pathtracing/mgpu/weightGpu0'
No setting was found for '/rtx/pathtracing/mgpu/printLog'
No setting was found for '/rtx/pathtracing/mgpu/printWeights'
No setting was found for '/rtx/pathtracing/mgpu/showCompositingPattern'
The log also contains:
MultiGPU Board : No
There is also a previous crash-reporter entry associated with the Multi-GPU launch command.
After stopping Movie Capture, all GPUs correctly return to idle:
GPU 0–3: 0% utilization, 0 MiB memory used
Questions
- Is RTX Path Tracing Multi-GPU supported for Movie Capture in Isaac Sim 6.0.0?
- Is
/rtx/pathtracing/mgpu/enabledthe correct setting for this workflow? - Is another setting required for Movie Capture or render-product capture?
- Should “Use render product to capture” be enabled for Multi-GPU path tracing?
- Does
Samples per pixel per iterationaffect GPU distribution, or only batching? - What exact settings are required to distribute path-tracing work across GPUs 0–3?
- What does this log message mean?
MultiGPU Board : No
- Does this Brev GPU topology support RTX Multi-GPU path tracing, or is a specific NVLink/multi-GPU board topology required?
- Why does Isaac Sim report all four GPUs as active and allocate memory on all four while only GPU 0 performs rendering?
- Are there known restrictions when running Isaac Sim in Docker through Vulkan and noVNC?
Any exact command-line flags, Kit settings, Python API calls, or supported Movie Capture configuration would be appreciated.
I am preserving the VM and logs and can provide the complete Kit log privately if needed.
Thank you.