Nvidia CUDA Ending Support for Older GPU Architectures: Maxwell, Pascal, Volta
If you're a coder working with Nvidia's CUDA platform, there's some important news for older graphics cards. The latest official notes for CUDA Toolkit 12.9 make it clear: the next major release will no longer support compute features for GPUs from the Maxwell, Pascal, and Volta architectures.
Now, fear not, this does not mean your previous gaming card will simply stop working for games. Nvidia will likely still release regular GeForce drivers for these GPUs for a little while longer. This change directly speaks to the existing CUDA development side – the software employed to create applications that leverage GPU compute muscle.
The End of an Era for CUDA Development on Certain Architectures
Nvidia's message is very clear this time around. While there were earlier hints, it is now clear that developers must start planning how to port their CUDA projects onto newer GPU architectures. The current CUDA 12.x series (and earlier ones) will still enable you to write applications for these Maxwell, Pascal, and Volta cards. However, the future major release (most likely CUDA 13.x) will impose some significant limitations.
What exactly will be lost. The ability to natively compile code for these older GPUs (offline compilation) and compatibility with CUDA-accelerated libraries will be lost. That means future versions of the CUDA compiler (nvcc) won't be able to produce machine code that these older cards can run. Similarly, popular libraries like cuBLAS and cuDNN won't be compatible with these architectures in their future versions.
"Maxwell, Pascal, and Volta architectures are feature-frozen without further changes on anything on the roadmaps. While these architectures will continue to be supported for building applications under the CUDA Toolkit 12.x series, offline compilation and library support will be removed in a future major version of the CUDA Toolkit. Migration planning to more recent architectures must be done by customers, as toolkits in the future will no longer be capable of targeting Maxwell, Pascal, and Volta GPUs."
- CUDA 12.9 Toolkit release notes
Nvidia hasn't released a clear date on when precisely this new major CUDA update will be out, nor whether there are how many minor updates that we're going to have within 12.9.x series. But take it for granted, cutting support for three big architectures simultaneously is a huge deal. The subsequent architecture in turn is Turing (employed in RTX 20-series cards), which is left alone for the time being and likely has some life left before facing the same fate.
A Look Back at the Affected Architectures
It's worth noting to remember the impact these architectures had:
- Maxwell: Released back in the early days of 2014, Maxwell propelled cards like the GTX 750 Ti and the GTX 900 series, which were highly sought after. It even made it into the original Nintendo Switch.
- Pascal: Pascal, in 2016, laid the foundation for legendary cards like the GTX 1080 Ti, Quadro workstation GPUs, and Tesla accelerators.
- Volta: Although consumers might not know Volta's name, it was an essential architecture that brought Nvidia's Tensor Cores in 2017. Volta-based GV100 chip was enormous and represented a big step by Nvidia on its path towards AI dominance and set the tone for upcoming architectures like Turing, Ampere, Hopper, and the latest Blackwell.
This shift is the natural part of technology advancement, but it's an undisputed call to action for developers to adapt and turn toward Nvidia's more recent GPU innovations for their future CUDA-led endeavors.