Karma XPU crash A6000

   976   7   0
User Avatar
Member
32 posts
Joined: Feb. 2016
Offline
Windows 10
GPU: Nvidia RTX A6000
GPU Driver: 551.52

Using Houdini 20.0.547/590 Apprentice. This is an animation from Mixamo, applying a material x shader. The shot renders fine on 4090 and 3080's. However on this A6000 it crashes. The log is not very helpful, even though I have increased the console log to 1024, I think apprentice limits the output to the default 128 or 256 maybe.

When I render a single frame to mplay, I get exit code 139. When I render in viewport, it crashes houdini completely.

The geo renders fine on the A6000 without shaders. If I render with ONLY material x nodes in the shader, it's fine. However, the shader uses the Karma round corners node into a mix. If I remove that node, the render works.

In viewport the render works but I can see the GPU is not being used, the Optix device is "compiling". Embree is 100%. Then after a moment it crashes. Could it be, once the GPU actually kicks in, it fails?

Any help would be appreciated. Would love to take advantage of this GPU.
User Avatar
Member
32 posts
Joined: Feb. 2016
Offline
Here's the shader network.

Also, Karma CPU is fine.

Edit: I stumbled across the XPU environment variables(ie, KARMA_XPU_DISABLE_EMBREE_DEVICE = 1 and KARMA_XPU_DISABLE_OPTIX_DEVICE = 1). I can now confirm that the Embree device is fine but the Optix device is the issue for some reason.
Edited by jarjarshaq - Feb. 16, 2024 15:21:41

Attachments:
shader.png (392.5 KB)

User Avatar
Staff
480 posts
Joined: May 2019
Offline
What happens if you try this environment variable?
KARMA_XPU_OPTIX_ENABLE_OPTIXIR=1

If that still does not work, what happens if you roll back to some 545+ driver? (eg 546.33?)
Please keep in mind, if you use this driver, then you'll need to set the above environment variable to 0
User Avatar
Member
32 posts
Joined: Feb. 2016
Offline
I can give this a try tomorrow and report back. Out of curiosity, what does this environment variable do? I don't see it listed in the docs.
User Avatar
Staff
480 posts
Joined: May 2019
Offline
jarjarshaq
I can give this a try tomorrow and report back

thanks

jarjarshaq
Out of curiosity, what does this environment variable do? I don't see it listed in the docs.

It makes XPU send code/instructions to Optix using the newer OptixIR format, rather than the traditional PTX format.

Its a developer switch that we use for debugging and would normally make no difference to a user. But in this case the PTX interpreter is crashing so it offers a temporary workaround. We don't document these internal/developer/debug switches because they change often, are somewhat meaningless to a user, and can sometimes cause instability if used incorrectly.
User Avatar
Member
11 posts
Joined: Oct. 2021
Offline
Hi,

Sorry I am barging in this topic, but I feel it pertains to me.
I am currently communicating with SideFX support (ticket #149148) in regards to a similar problem I have - Karma XPU crashing (win 11, RTX 4090, latest studio driver) as soon as I enable and start rendering any geometry lights in a scene.

I tried everything that the support ticket was telling me to do but nothing was helping. Really frustrated I cannot finish this project.

I googled and found this thread, and tried the KARMA_XPU_OPTIX_ENABLE_OPTIXIR=1 variable, and I can say that Houdini is finally NOT crashing as soon as I enable geometry lights woo!!!

Is this a permanent solution, or just a debugging tool as you said @brianns?

Any help is appreciated.
Edited by chilby - Feb. 20, 2024 06:39:23
User Avatar
Member
32 posts
Joined: Feb. 2016
Offline
brians
KARMA_XPU_OPTIX_ENABLE_OPTIXIR=1
This now works with the environment variable set. So I'll leave this set this way till Nvidia release a driver update that fixes this issue.
User Avatar
Staff
480 posts
Joined: May 2019
Offline
chilby
Is this a permanent solution, or just a debugging tool as you said @brianns?

You should be fine leaving this enabled (for any 550+ driver) until you hear otherwise.
We'll reply on this thread once we hear back from NVidia in the coming weeks.
  • Quick Links