Hey guys First of all happy new year everyone!

I recently noticed that ramps like the Karma Ramp Constant and the Karma Ramp Parameter seem to be completely broken in Karma XPU rendering. It's very easy to reproduce; Take any roughness texture map, feed it into a ramp to remap values and then into a mtlX standard surface or openPBR standard shader. While the look changes initially compared to not having the ramp plugged in, any further tweaking of values is not reflected visually. Switch over to CPU however and everything is working as expected. So basically right now there is now way of using ramps in shaders if you wanna render with XPU which is a major problem imho. I can't rule out that it also might be an Nvidia driver issue. I've updated to the latest studio drivers two or three days ago, so maybe that might also be the issue? Maybe someone who hasn't updated yet can cofirm the issue? I also always precompile shaders after updating my driver(new feature in H21), maybe something could've gone wrong there?
The only way around that I found so far is importing the texture maps into COPS, do the remap there and then pull the result back into the VOP network with the 'op:' syntax. While this works generally there's one major problem; op: doesn't work for UDIM textures (would be really nice if it did btw) so the only real option right now is baking out remapped texture maps from COPS and read them in with mtlx image nodes in the shader which is painstakingly slow and horrible for iteration.

Looking forward to hearing back from you and if I'm alone with this problem or if it's actually a bug.

I tested this on 21.0.585, .588, .590 daily builds and also on .559 production.
System is Windows 11 25H2, Nvidia RTX 4090 Studio driver v591.74, 128GB DDR5 RAM

It seem the NVidia 590 range of drivers have a bug.
What happens if you roll back to a 580 driver? (eg 581.57)

brians
It seem the NVidia 590 range of drivers have a bug.
What happens if you roll back to a 580 driver? (eg 581.57)

Do you have any source for that or anything to read up on? What else does this bug affect? That would kinda influence my decision if rolling back drivers and recompiling all shaders again for hours is worth it to me.

> Do you have any source for that

Communication with NVIDIA engineers have confirmed it's a 590 bug, which should be fixed in 595.

> What else does this bug affect?

Not sure yet, it seems to be only affecting the kma_ramp nodes for now. A compiler glitch it seems.

> That would kinda influence my decision if rolling back drivers and recompiling all shaders again for hours is worth it to me.

I've committed a workaround for the issue to 21.0.594 (which will make its way into the next production build, hopefully this week).

But this change will trigger the need for a recompile anyway, so it's up to you if you...
- stay on 590, upgrade to 21.0.594(or later)
- roll back to 581.57, and stay on your current Houdini version.
(either way a recompile will be needed sadly)

Thanks so much for the helpful update. Guess I'm gonna do the little precompile dance with the new daily then and then in a couple days once again when the new driver releases haha. At least it doesn't take thaaaat long on a 4090. Kinda ironic that that a driver with the version number "590" has some issues 😅 BTW: Do you think it would make sense it's be an option to offer the cache as an optional download with Houdini? I know that everyone has different GPUs and drivers and all but maybe just for a handful of cards like the 80s and 90s of the current and last gen and only for the current driver that's available at the time? I believe the cache is not that big, around 6GB. I'm sure many would appreciate that.

> Do you think it would make sense it's be an option to offer the cache as an optional download with Houdini?

We'd really like to get to this point, but at the moment NVidia own the cache and its a black box to us. And its heavily specialized to CPU, machine makeup, etc... (ie not just GPU + driver). So there is nothing we can do today, but we are working with NVidia to address this in future releases, so watch this space.

Good to know people are making use of the pre-compile feature in anycase