CUDA 13 ends support for Pascal and Volta - what it means for a used Tesla for AI

If you are looking for an inexpensive second-hand card for AI, you have probably already come across alarming headlines: "NVIDIA ends CUDA support for Pascal and Volta". It sounds like a death sentence for the Tesla P40 or V100, the most popular used cards for local models. The short answer: the cards do not stop working, but how long the newest software versions will support them is changing. Below I break down exactly what was announced, which cards it affects and how it really affects a purchase, with references to the sources so that it can be checked.

What exactly NVIDIA announced

This is not a rumour or an interpretation: it is in the official CUDA release notes. The process took place in two steps:

Step 1: "feature-complete" status (CUDA 12.x). NVIDIA marked the Maxwell, Pascal and Volta architectures as "feature-complete": no further improvements or new features. The CUDA 12.x series still allowed applications to be built for these cards, but with the announcement that offline compilation and library support would disappear in the next major release (CUDA 12.9 release notes; also reported by Tom's Hardware) [2][7].

Step 2: removal of support (CUDA 13.0). In the CUDA Toolkit 13.0 release, offline compilation and library support for these architectures was removed, specifically for cards with a compute capability earlier than Turing [3][1]. In practice this means that future compilers (nvcc) will not generate machine code for these GPUs, and CUDA-accelerated libraries such as cuBLAS and cuDNN will no longer support them in new versions [1].

The key distinction that gets lost in the headlines: this is a removal of support on the compute/toolkit side, not an immediate "switching off" of the cards. Drivers are a separate axis (more on that shortly).

Which cards it affects

The dividing line is Turing. Everything older (Maxwell, Pascal, Volta) drops out of the new toolkits; Turing and newer (including Ampere) remain [1][3]. Translated into the accelerators that are actually bought used for AI:

Card	Architecture	Compute capability	Status in CUDA 13+
Tesla P4	Pascal	6.1	⚠️ compilation/library support removed
Tesla P40	Pascal	6.1	⚠️ removed
Tesla P100	Pascal	6.0	⚠️ removed
Tesla V100	Volta	7.0	⚠️ removed
NVIDIA A2	Ampere	8.6	✅ supported

The compute capability of the cards is confirmed by NVIDIA's data [4]. The conclusion is unequivocal: of the five cards in the table, four (all of Pascal + Volta) have lost support in the new toolkits, and from our current range the only one that remains in the mainstream of support is Ampere, that is, the A2. This is not the limit of what is possible: other cards based on the Turing, Ampere and Ada architectures also retain support in CUDA 13, and those we do not happen to have in stock we can source to order.

Driver ≠ toolkit: two different axes that are easy to confuse

This is the most common misunderstanding around the matter, so it is worth separating them.

The toolkit and libraries (compute). This is the core of the change described above: new versions of CUDA (13+), cuBLAS, cuDNN and the frameworks built on them will stop including code for Pascal/Volta. The ecosystem has already put this into effect: PyTorch 2.8 removed Maxwell and Pascal from the CUDA 12.8/12.9 builds, and PyTorch 2.11 removed Volta from them as well; CUDA 12.6 wheels remain the prebuilt option for these cards [5][12]. PyTorch 2.14 is the last release with prebuilt CUDA wheels for these cards; running newer PyTorch versions on these GPUs from 2.15 onward requires a source build against CUDA 12.x [9]. Some tools have moved on even earlier: current vLLM's NVIDIA backend requires compute capability 7.5 or higher, and the main upstream CUDA implementation of FlashAttention-2 and native BF16 arithmetic require Ampere or newer, so the A2 supports them and the P40, P100 and V100 do not (software-emulated BF16 and other attention implementations are a separate matter) [10][11].

Drivers. This is a separate cycle. For GeForce cards based on these architectures, NVIDIA has announced that after the last Game Ready driver of October 2025 they will move to quarterly security updates for a further three years, until October 2028 [6]. Note: this is an announcement about gaming drivers for GeForce. For data-centre (Tesla) cards NVIDIA gives a date too: the R580 branch, released alongside CUDA 13.0, is the last driver branch that supports GPUs below compute capability 7.5, and as a long-term support branch it is maintained until mid-2028 [8]. Branches 590 and later no longer support the P4, P40, P100 or V100 (the driver documentation moves them to the 580.xx legacy list), so on these cards you pin the driver to the 580 branch as well [13].

Does a used Tesla suddenly become useless? No, and this is not downplaying it

Removing support in new software is not the same as "the card has died". Three facts help clarify this:

1. Existing software keeps working. CUDA 12.x, and the libraries and frameworks built for these architectures, do not disappear: you can use them for as long as you maintain compatible versions. llama.cpp built with CUDA 12.x (for example its cuda-12.4 release builds for Windows and cuda-12.8 for Linux), Ollama (on Pascal it needs driver 570 or newer, so the 570-580 window) and PyTorch up to 2.14 with CUDA 12.6 wheels will run on the P40 or V100 just as they do today; builds made with CUDA 13 will not.
2. The change concerns the "newest" releases, not the whole stack. The problem arises when you always want the freshest version of a library built with the newest toolkit: that is what will, over time, stop including code for these cards. It is a question of update strategy, not of the hardware's condition.
3. This is a normal life cycle, not a failure. Pascal debuted in 2016 and Volta in 2017: after ~8–9 years, ending up on the phase-out list is predictable, not surprising.

In other words: it is a reason to make an informed purchase, not to avoid buying.

What this means for a purchase (analysis)

What follows is my own interpretation of the facts gathered above; I am keeping it clearly separate from the data.

Buying "for now and a few years", on a low budget, for local LLMs? Pascal (especially the P40 24 GB) and the V100 still make sense: they give the most compute and VRAM for the money. The condition: accept that this is hardware with a horizon of a few years, and pin the versions of your software (CUDA toolkit 12.x, the 580 driver branch and PyTorch up to 2.14 installed from the CUDA 12.6 wheels) instead of blindly updating to the newest.

Buying "for years", for an institution, or wanting peace of mind with updates? The guidance is now unequivocal: Ampere, that is, the A2: it is the only card in our current range that remains in the mainstream of CUDA support. You pay for a newer architecture and accept lower raw performance in return for the longest support horizon. If you want something newer or more powerful that is still supported long term (another card from the Turing, Ampere or Ada architecture), we can source it to order, with a quotation on request.

Doing HPC/FP64 or heavier training? Here the decision is more complex and goes beyond CUDA 13 itself: FP64, bandwidth and VRAM all matter. That is a topic for a separate article; when buying, start in any case with the checklist for used data-centre hardware.

One thing is certain whatever the model: with hardware from the secondary market, the condition of the unit (a load test) matters more today than a "support horizon" on paper.

Conclusions

CUDA 13 is not the "death" of Pascal and Volta, only the end of their presence in the newest compilation toolchains, with R580 as their last driver branch, supported until mid-2028 [1][3][8]. For someone buying a used card for AI, this means a simple decision rule: the A2 (Ampere) for years; the P40/V100 for now and a few years, with deliberate version pinning. The hardware does not disappear: only the window of "always the newest" software shrinks, and that is a trade-off that, at the price of these cards, still pays off for many people.

Sources

1. NVIDIA: CUDA Toolkit Release Notes (feature-complete / removal of support for Maxwell/Pascal/Volta): docs.nvidia.com/cuda/cuda-toolkit-release-notes/
2. Tom's Hardware: "Nvidia to drop CUDA support for Maxwell, Pascal, and Volta GPUs with the next major Toolkit release" (based on the CUDA 12.9 notes): tomshardware.com/pc-components/gpus/nvidia-to-drop-cuda-support-for-maxwell-pascal-and-volta-gpus-with-the-next-major-toolkit-release
3. NVIDIA: CUDA Toolkit 13.0 Release Notes (removal of support for compute capabilities earlier than Turing): docs.nvidia.com/cuda/archive/13.0.1/cuda-toolkit-release-notes/
4. NVIDIA: CUDA GPUs / Compute Capability (Tesla P4/P40 = 6.1, P100 = 6.0, V100 = 7.0, A2 = 8.6): developer.nvidia.com/cuda/gpus
5. PyTorch: GitHub issue #157517: "[release 2.8-2.9] Delete support for Maxwell, Pascal, and Volta architectures for CUDA 12.8 and 12.9 builds" (closed as completed, 16 January 2026): github.com/pytorch/pytorch/issues/157517
6. Tom's Hardware: "Nvidia confirms end of Game Ready driver support for Maxwell and Pascal GPUs: … optimized drivers through October 2025" (+ quarterly security updates until October 2028): tomshardware.com/pc-components/gpus/nvidia-confirms-end-of-game-ready-driver-support-for-maxwell-and-pascal-gpus-affected-products-will-get-optimized-drivers-through-october-2025
7. NVIDIA: CUDA Toolkit 12.x Release Notes (archive), "feature-complete" status for Maxwell/Pascal/Volta: docs.nvidia.com/cuda/archive/12.9.0/cuda-toolkit-release-notes/ (for the CUDA 12.x point, refer to the archived version, not to the current notes, which always point to the newest release)
8. NVIDIA Technical Blog: "Navigating GPU Architecture Support: A Guide for NVIDIA CUDA Developers" (4 August 2025), R580 as the last driver branch for compute capability below 7.5, supported until mid-2028: developer.nvidia.com/blog/navigating-gpu-architecture-support-a-guide-for-nvidia-cuda-developers/
9. PyTorch dev-discuss: "Notice: CUDA 12.6 wheels will no longer be published from PyTorch 2.15 (drops Maxwell, Pascal, Volta)": dev-discuss.pytorch.org/t/notice-cuda-12-6-wheels-will-no-longer-be-published-from-pytorch-2-15-drops-maxwell-pascal-volta/3432
10. vLLM: GPU installation (compute capability 7.5 or higher): docs.vllm.ai/en/latest/getting_started/installation/gpu.html
11. Dao-AILab/flash-attention README (Ampere, Ada, Hopper; native BF16 requires Ampere or newer): github.com/Dao-AILab/flash-attention
12. PyTorch 2.11.0 release notes: CUDA 12.8 and 12.9 builds without Volta, CUDA 12.6 builds for Volta: github.com/pytorch/pytorch/releases/tag/v2.11.0
13. NVIDIA Linux driver 590.44.01 README, appendix A "Supported NVIDIA GPU Products": Tesla P4, P40, P100 and V100 on the 580.xx legacy list: download.nvidia.com/XFree86/Linux-x86_64/590.44.01/README/supportedchips.html

---

We run this article on the Agile Store blog, where we deal in used NVIDIA Tesla accelerators: whichever architecture you choose, with a used card ask first for a load test (gpu-burn), and choose your software versions to match the card's compute capability.