[BUILD] CUDA 13 drops support for pre-Turing architectures

[jfpanisset]

While trying to build OSL 1.15.5.0 with LLVM 22.1.7 and CUDA 13.3.0 I got the following build failure:

nvcc fatal   : Unsupported gpu architecture 'sm_60'

when trying to build:

src/testshade/optix_grid_renderer.ptx
src/testrender/optix_raytracer.ptx

CMakeLists.txt has:

set (CUDA_TARGET_ARCH "sm_60" CACHE STRING "CUDA GPU architecture (e.g. sm_50)")

which calls for Pascal architecture. I updated my Conan recipe to set that to "sm_75" (Turing) when detecting CUDA 13 or newer, I don't know if OSL makes any claims about CUDA 13 compatibility yet, but it might be good to have CMakeLists.txt adjust the baseline GPU architecture based on the detected CUDA version?

[lgritz]

Is there a succinct chart somewhere that shows the architectures supported for each CUDA version?

@ee33 @timgrant @chellmuth What's a reasonable min/max (a) CUDA version (b) SM architecture (c) OptiX version for OSL to support in 2026? Maybe it's time to advance the minimums of these that we support, it's probably been a while. And certainly we shouldn't have the minimum of one of of those be so old that the others at their minimums can't support it, that wouldn't be helpful to anyone.

[jfpanisset]

There doesn't seem to be an authoritative (or even better machine readable) list of "minimum architecture supported by each CUDA version", apparently this user maintained StackOverflow table may be the closest thing:

https://stackoverflow.com/questions/28932864/which-compute-capability-is-supported-by-which-cuda-versions

[lgritz]

Ideally, I just want an oracle to tell me the set of { min CUDA toolkit, min OptiX, min CUDA target arch, reasonable default CUDA target arch } to use for the next set of releases.

[jfpanisset]

Anecdotally I would say that most facilities should have cycled out Pascal generation GPUs, since those are 2016 vintage and starting to lack a lot of the more modern functionality expected by current DCCs. Also Pascal GPUs are last supported by the legacy 580 NVIDIA driver branch (current branch is 595 I believe).

Unfortunately "how far back should you go for GPU / GPU driver support" is a question that has no clear answer:you don't want to orphan a bunch of hardware for no good reason, but you also don't want to skip using new functionality to keep supporting hardware that isn't relevant in a studio.

VFX Platform explicitly does not take a position on minimum GPU generations / capabilities or driver versions.

Best I could see would be some kind of informal annual survey via ASWF or StudioSysadmins to get an idea of what people are running out there.

[tgrant-nv]

There was a blog post about this a while back that might be helpful. Navigating GPU Architecture Support: A Guide for NVIDIA CUDA Developers

Here's the summary from that post:

If you’re writing applications that support GPUs of CC earlier than 7.5, then you’ll want to install and remain on the following:

• NVIDIA Driver: Remain on branch 580. This provides support for 3 years, through sometime in 2028.
• CUDA Toolkit: Remain on 12.9. This provides support for offline compilation for the pre CC 7.5 GPU architectures.

In summary, the upcoming CUDA 13.0 release will remove support for offline compilation for GPU architectures before compute capability 7.5. Developers who want to continue building applications for these GPUs should continue using the NVIDIA driver branch 580 and CUDA Toolkit 12.9 to maintain support for these older architectures.

The "older architectures" are the pre-Turing GPUs: Maxwell, Pascal, Volta.

[lgritz]

Do you think it's time to bump the default for OSL compilation to 75? People can still lower if they know they have the right cards and drivers.

For the kind of code that OSL is likely to generate, how important is it for people to change that default of 75 to something newer that represents the hardware they actually have? Am I wrong to assume that the driver is probably pretty good about taking the PTX and turning it into whatever SASS is right for the actual GPU they have?