It is said CUDA13 stops supporting Volta arch, but I found it’s able to run few binaries built with gcc15(or clang20)+cuda13 on V100 device stable without issues. Is it safe to do so or could have undefined behaviour? or just specific aspects of CUDA calls could have issues?
Hi — a few thoughts in case they’re useful.
You’re right that some CUDA 13 binaries do run on V100, but it’s probably less “supported” and more
“happens to work because of PTX forward-compat.” When a binary ships with PTX alongside (or instead of)
SASS, the driver can JIT it at launch time for whatever architecture is actually installed, including
Volta. So simple kernels built with gcc15+cuda13 will often just work on V100 via JIT, even though
Volta was dropped from the toolkit.
Where it gets risky is the libraries. CUDA 13 removed Maxwell/Pascal/Volta from the default support
matrix, which means cuBLAS, cuDNN, cuSPARSE, cuFFT, Thrust, etc. built for 13 may have dropped the
sm_70 kernel variants entirely. So:
- Hand-written kernels: usually fine via PTX JIT.
- Anything that calls into CUDA-X libraries: may work for common paths and then hit “no kernel image is
available for execution on the device” (or worse, silently degraded results) when it trips over a
missing sm_70 specialization. - Numerical intrinsics / Tensor Core paths: some got Volta-specific implementations removed, which can
change behavior silently.
So the “stable on a few binaries” observation is probably real but fragile — which binaries, which ops,
which input shapes all matter, and it can break on a library or driver update.
If you need CUDA 13-era features on V100, one workaround some people use is mixing: build with
–generate-code=arch=compute_70,code=compute_70 so PTX is embedded, and stay on older cuBLAS/cuDNN that
still ship sm_70 kernels. Otherwise CUDA 12.x is the last branch with first-class Volta support and is
probably the safer target for anything V100 needs to run reliably.
Some references:
- CUDA 13 release notes (deprecated arch list):
CUDA Toolkit 13.2 - Release Notes — Release Notes 13.2 documentation - CUDA compatibility / forward-compat explainer: CUDA Compatibility — CUDA Compatibility
- GPU compute capability table (V100 = sm_70): CUDA GPU Compute Capability | NVIDIA Developer
Just passing by so take this as a lead rather than a definitive answer — not sure I’ll be around to
follow up, but hope some of it’s useful.
Thank you for your reply, it’s very useful.