[Bug] Isaac Sim 4.5.0.0 crashes on RTX PRO 5000 Blackwell (CC 12.0) — SimulationApp segfault in _wait_for_viewport

[omkathalkar]

Description

Environment

• OS: Ubuntu 24.04.4 LTS (kernel 6.8.0-124-generic)
• NVIDIA Driver: 595.71.05
• CUDA Toolkit: 12.6 (nvcc V12.6.85)
• Isaac Sim: 4.5.0.0 (pip install, isaacsim[all]==4.5.0.0)
• Conda env: Python 3.10
• GPU: NVIDIA RTX PRO 5000 Blackwell (48GB, CUDA CC 12.0, PCI 02:00.0)
• Second GPU: NVIDIA GeForce RTX 4060 Ti (16GB, PCI 01:00.0)

---

Issue

SimulationApp segfaults during initialization when the RTX PRO 5000 Blackwell is present. The crash is consistent across headless=True, no_window=True, and all renderer options tested (RaytracedLighting, PathTracing).

Minimal repro:
from isaacsim import SimulationApp
app = SimulationApp({"headless": True, "no_window": True})
print("SimulationApp: OK")
app.close()

---

Root cause (from logs)

iRay explicitly reports Blackwell as unsupported:
[Warning] [rtx.neuraylib.plugin] [IRAY:RENDER] 1.1 IRAY rend warn :
GPU 1 (NVIDIA RTX PRO 5000 Blackwell) with CUDA compute capability 12.0
is unsupported by this version of iray photoreal.
Please look for an update with your software provider.

The RTX scene database also overflows on the 48GB VRAM:
[Warning] [rtx.scenedb.plugin] SceneDbContext : TLAS limit buffer size 7512601600
[Warning] [rtx.scenedb.plugin] SceneDbContext : TLAS limit : valid true, within: false

The resulting segfault crash trace:
Fatal Python error: Segmentation fault

Thread (most recent call first):
File "simulation_app.py", line 520 in _wait_for_viewport
File "simulation_app.py", line 270 in init

With detailed async trace showing the rendering crash propagating through the Kit event loop:
File "app_iface.py", line 51 in _next_update_async
File "material_library.py", line 1656 in get_mdl_list_async
File "material_extensions.py", line 36 in add_menus
File "async_engine.py", line 197 in run_once
File "simulation_app.py", line 520 in _wait_for_viewport
File "simulation_app.py", line 270 in init

---

Workaround

Forcing rendering onto the RTX 4060 Ti (Omniverse device index 1) with multi-GPU disabled works:
app = SimulationApp({
"headless": True,
"no_window": True,
"active_gpu": 1,
"multi_gpu": False,
"max_gpu_count": 1
})

This is not ideal — the RTX PRO 5000 Blackwell is the primary workstation GPU and should be used for Isaac Sim rendering.

---

Request

Please update iRay and the RTX scene database in Isaac Sim to support CUDA compute capability 12.0 (Blackwell architecture). Confirmed working in Isaac Sim 4.5.0.0: 4.5.0.0 is the latest available on PyPI — no newer version to upgrade to.

Isaac Sim version

4.5.0.0

Operating System (OS)

Ubuntu 24.04.4 LTS

GPU Name

NVIDIA RTX PRO 5000 Blackwell (48GB, CUDA CC 12.0, PCI 02:00.0)

GPU Driver and CUDA versions

595.71.05

Logs

No response

Additional information

No response

[vick-yu]

Thanks for the detailed report and the clear isolation of the iRay warning -- that saves a lot of debugging time.

What's Happening

Isaac Sim 4.5.0 ships Kit SDK 106.5, which predates Blackwell architecture (CUDA compute capability 12.0 / sm_120) support in both the iRay renderer and the RTX scene database. The segfault in _wait_for_viewport is a downstream consequence of the renderer failing to initialize on an unsupported GPU architecture. This is not something that can be patched into the 4.5.0 release -- the underlying Kit SDK and renderer binaries need the Blackwell support that was added in newer versions.

Additionally, your driver 595.71.05 is from the R595 New Feature Branch (NFB), which is not validated for any Isaac Sim release. NFB drivers frequently cause rtx.scenedb.plugin and renderer crashes even on supported GPU architectures (we've seen this pattern across several Blackwell reports: #619, #643, #651).

Recommended Steps

1. Upgrade to Isaac Sim 6.0.0 (current GA release), which has native Blackwell support -- the RTX PRO 6000 Blackwell is even listed in the system requirements as the "Ideal" GPU:

   pip install isaacsim[all]==6.0.0.1

   Note: Isaac Sim 6.0.0 requires Python 3.12 (4.5.0 used Python 3.10), so you will need to recreate your conda environment. See the migration guide for API changes.

2. Switch to the R580 production branch driver (580.95.05+ recommended for Linux):

   • Download: https://www.nvidia.com/en-us/drivers/details/254666/
   • The R595 NFB driver is not validated and is a known source of renderer crashes with Omniverse applications.

3. If you need to stay on 4.5.0 temporarily, your workaround of forcing rendering onto the RTX 4060 Ti with active_gpu / multi_gpu / max_gpu_count is the right approach. But 4.5.0 will never support CC 12.0 -- the only path forward for the RTX PRO 5000 is upgrading.

Why 4.5.0 Shows as "Latest" on PyPI

Isaac Sim 5.x and 6.0.0 are published to PyPI, but each major version pins a specific Python version: 4.x requires Python 3.10, 5.x requires Python 3.11, and 6.0.0 requires Python 3.12. Since you're using a Python 3.10 conda environment, pip's resolver filters out anything with requires_python==3.11.* or ==3.12.*, making 4.5.0 appear to be the latest compatible version. Running pip install isaacsim[all]==6.0.0.1 in a fresh Python 3.12 environment will pick up the current GA release.

Let us know if the crash persists after upgrading to 6.0.0 with the R580 driver -- that combination is validated for your GPU.