Hi,

I'm hitting a reproducible Karma XPU failure in Solaris viewport (IPR) rendering and wanted to check if anyone else is seeing this on newer NVIDIA drivers, or if SideFX is already aware.

**System:**
- Houdini 21.0.729
- CachyOS (Arch-based), kernel 7.1.3-2-cachyos
- NVIDIA driver 610.43.03, open kernel modules (nvidia-open)
- 2x RTX 4090 (one driving the display, one headless/compute only)
- Wayland (KDE Plasma)

**Symptom:**
Shortly after First Pixel in a Solaris viewport render (~2 seconds in), both OptiX devices fail simultaneously:

```
KarmaXPU: device Type:Optix ID:1 Name:"NVIDIA GeForce RTX 4090" has registered a critical error "cudaErrorIllegalAddress", so will now stop functioning. Future error messages will be suppressed
KarmaXPU: device Type:Optix ID:0 Name:"NVIDIA GeForce RTX 4090" has registered a critical error "cudaErrorIllegalAddress", so will now stop functioning.
KarmaXPU: Aborting render - no working device
```

Kernel log at the same moment shows an MMU fault — a virtual read at address 0x0 (null read) inside the houdini-bin process:

```
NVRM: Xid (PCI:0000:34:00): 31, pid=8686, name=houdini-bin, channel 0x00000019, intr 00000000. MMU Fault: ENGINE GRAPHICS GPC4 GPCCLIENT_T1_1 faulted @ 0x0_00000000. Fault is of type FAULT_PDE ACCESS_TYPE_VIRT_READ
```

As expected with this error, restarting XPU from the viewport doesn't recover the devices — only a full Houdini restart does. The GPUs themselves sometimes stay alive and visible in nvidia-smi, but sometimes only full hard reset helps.

**Possibly relevant details:**
- The console shows several "Displacement shader ... doesn't modify P (disabling)" warnings right before the crash, so displacement may be involved.
- The scene uses MaterialX materials with texture maps (basecolor/roughness/normal/metallic), an HDRI dome, and an OptiX denoiser.
- Windows on the same machine renders the same kind of workload fine, so this looks Linux-driver-specific.
- This feels similar in character to the driver 565 issue from 2024 (https://www.sidefx.com/forum/topic/98442/) that also produced cudaErrorIllegalAddress under XPU/OptiX and was worked around in a Houdini build.

**Question:** Is driver 610.x validated against Houdini 21.0, and is there a known issue with it (or with nvidia-open modules specifically)? Happy to provide a repro scene, nvidia-bug-report.sh output, and exact repro steps to support if useful.

Thanks!

Another note, it happens on windows as well, so I guess it is drivers incompatibility with XPU and only road to take is roll back drivers to old versions I assume?

sorry but CachyOS (ie Arch-Linux) is not one of our supported Linux distributions
https://www.sidefx.com/Support/system-requirements/ [www.sidefx.com]

Mirko Jankovic
Another note, it happens on windows as well, so I guess it is drivers incompatibility with XPU and only road to take is roll back drivers to old versions I assume?

ah, sorry I thought you said windows was working fine

Mirko Jankovic
**Question:** Is driver 610.x validated against Houdini 21.0, and is there a known issue with it (or with nvidia-open modules specifically)? Happy to provide a repro scene, nvidia-bug-report.sh output, and exact repro steps to support if useful.

We're not aware of any issue, but have not done extensive testing yet.
Repro scene + exact repro steps would be great thanks. As simple scene as possible is ideal.

thanks

I'm seeing the same behavior even in the latest scene from your crab rigging tutorial, which was released recently.
https://www.sidefx.com/tutorials/rig-animate-a-stylized-crab/ [www.sidefx.com]
CH6_rendering_01.hip
How to reproduce:
Just turn on Karma XPU in Solaris and start working — timeline changes, camera movements, etc. At some point, the 2nd GPU just drops.
BUT, I've now tested the same thing on two other machines, both with dual 4090s on Windows with the latest NVIDIA gaming drivers, and everything seems fine there.
I did more testing on my machine and replaced the riser cable (the GPUs are on risers due to the motherboard's PCIe slot configuration and the thickness of the cards). Nothing helped, so for the time being I guess I'll just use a single GPU in this system. It could be the motherboard (ASRock Taichi X570) or something else — I'm not sure, but it doesn't seem like I'll be able to pin down the exact issue.
What's odd is that it starts out working fine, runs okay for a while, and then the GPU drops off the PCI bus. Rendering on the Deadline farm can go much longer, but past a certain point it can happen again, so... I guess I'll chalk it up to gremlins and look into a new CPU/motherboard combo for these two GPUs.
Sorry for the noise — the issue is probably too specific to my setup to be reproducible on different machines.

Mirko Jankovic
Sorry for the noise — the issue is probably too specific to my setup to be reproducible on different machines.

No worries, its good to raise this kind of issue regardless.
One thing, what happens if you roll back the driver? to perhaps a 595 range?

Here's a cleaned-up version, keeping your casual forum tone:

Hmm, seems like the problems are gone in H22!
I had my GPU out... but figured, wth, let me try it. Because my H21 was freezing in the viewport as well, even with a single GPU. Not a full freeze, but so slow that it took a couple of minutes to register a click or a move... only X and save-and-close worked.
In Solaris, the 2nd GPU was dropping out. Looking at the logs, it seems like they both crashed, but the 1st one was able to recover.
Now... I've been testing H22 for hours — not a single glitch. CachyOS, latest NVIDIA drivers... all fine.
Figured I'd try again with both GPUs, and everything works, even Solaris! Spinning and rendering for over 15 minutes now... all fine. I'll of course test more, but before, it couldn't last longer than a couple of minutes...
So whatever it is... I'll take it and hope it won't break again!