The short answer: as CUDA 13 drops the Maxwell, Pascal (P4, P40, P100) and Volta (V100) architectures (CUDA 12.9 is the last toolkit that compiles for them), cards based on Turing (T4, sourced to order) and Ampere (A2 16GB, available from stock) remain in official support. From stock, that means the A2; the T4 we supply to order. And if you need something more powerful, we can source any accelerator still supported by CUDA 13 to order: the A100, A10, A30, A40, L4, L40 or H100. This does not mean that Pascal and Volta suddenly stop working: they still compute on CUDA 12.x. But if you care about the newest toolchain and a longer compatibility horizon, Turing/Ampere (and newer) have the advantage over them. A caveat straight away: "supported by CUDA 13" is not the same as a guarantee of many years of support, nor a promise that every future AI library will support a given card; more on that below.
I sell used Teslas, and I keep seeing waves of panic come back: "since CUDA is dropping Pascal and Volta, the P40 and V100 are e-waste now, right?". No, that is not true, and I do not want you to make a decision based on fear. I want to show you where the line really runs: when it is worth paying extra for a newer architecture, and when a cheaper P40 still does the job perfectly well. Honestly, with the drawbacks on both sides.
The key points to start with
- Fact: the T4 is a Turing card and the A2 is Ampere. Both remain in official CUDA Toolkit 13.x support while CUDA 13.0 removed Maxwell, Pascal and Volta.
- Also a fact: Pascal (P4/P40/P100) and Volta (V100) cards continue to work on CUDA 12.x and in llama.cpp; they do not turn into bricks. Details in a separate article on CUDA 13.
- A2 16GB (Ampere, compute capability 8.6): 16 GB of VRAM, configurable TDP of 40–60 W, low-profile form factor, BF16 support, hardware AV1 decoding. Ideal for dense packing and modern video.
- T4 16GB (Turing, compute capability 7.5): 16 GB of VRAM, 70 W, hardware video encoder/decoder. A proven workhorse for inference and transcoding, without hardware BF16 on the Tensor Cores and without AV1 decoding.
- A2 ≠ T4 in terms of being "future-proof": of the two, the A2 is the more future-proof design (a newer architecture, BF16, AV1), but the T4 is the faster card: 65 vs 18 TFLOPS of FP16 Tensor throughput, 8.1 vs 4.5 TFLOPS FP32 and 300 vs 200 GB/s of memory bandwidth. The A2 trades speed for low power and newer features; the T4 trades the newer features for throughput.
- Consider the A2/T4 when compatibility with a newer toolchain matters to you or you need newer codecs; the A2 can make sense for light fine-tuning, but 16 GB of VRAM severely limits the range of applications. Stay with the P40/V100 when you mainly do LLM inference in GGUF and the price per gigabyte of VRAM matters.
- Need more than a T4? We can source to order any accelerator supported by CUDA 13, from the A2/T4 to the A100, L4, L40 or H100. Write to us and we will help you choose for your budget and application.
What exactly CUDA 13 changes
The core of the matter is simple, and it is worth demystifying. An architecture being "dropped" in a new version of CUDA does not mean that the card stops working overnight. It means that the newest toolkit (CUDA 13) will no longer build or support code for these older chips. Pascal and Volta are frozen on the CUDA 12.x branch, and there they still compile, run and compute.
A lack of support in CUDA 13 does not automatically mean a lack of support for applications. Above all, it means that new code will no longer be developed and built for these architectures within CUDA 13. That is the difference between the end of support in a new toolchain and a situation in which the GPU stops working.
For most inference applications, this difference will not matter for a long time yet. llama.cpp, the popular runtimes for models in GGUF format and many ready-made stacks work very well on CUDA 12.x. Your P40 or V100 will not wake up tomorrow as e-waste; it will wake up as a card on an older but fully functional stack.
The difference only becomes significant when you want the newest libraries that require CUDA 13, or when you plan to keep the same hardware for many years and do not want to hit a wall of lost support halfway through that period. Then Turing and Ampere give you a longer runway.
An important distinction: CUDA Toolkit ≠ driver ≠ AI library. Support for an architecture in the CUDA Toolkit does not automatically mean that every new version of PyTorch, TensorFlow, FlashAttention or xFormers will support a given card. Individual libraries can raise the required compute capability independently of the toolkit itself, and Turing (7.5) is more exposed here than Ampere (8.6). "Supported by CUDA 13" gives you more freedom to use the new stack, but it does not guarantee that the newest version of every framework will run.
Table: what stays and what goes
| Card | Architecture (cc) | Future in the CUDA Toolkit | VRAM | Power draw | NVENC encoder | AV1 decoding |
|---|---|---|---|---|---|---|
| A2 16GB | Ampere (8.6) | Supported (remains) | 16 GB | 40–60 W (configurable) | Yes | Yes |
| T4 (to order) | Turing (7.5) | Supported (remains) | 16 GB | 70 W | Yes | No |
| P40 24GB | Pascal (6.1) | Removed in 13.0 → CUDA 12.x | 24 GB | higher TDP | Yes | No |
| V100 32GB | Volta (7.0) | Removed in 13.0 → CUDA 12.x | 32 GB | higher TDP | Yes | No |
| P4 8GB | Pascal (6.1) | Removed in 13.0 → CUDA 12.x | 8 GB | low TDP | Yes | No |
| P100 16GB | Pascal (6.0) | Removed in 13.0 → CUDA 12.x | 16 GB | higher TDP | Yes | No |
You will find the specific power figures for the Pascal/Volta cards on the product pages; what matters here is the pattern: the newer architectures (the A2 in particular) are much more energy-efficient, and the A2 is the only card in this comparison that decodes AV1 in hardware.
The second line: drivers, not just CUDA
The CUDA Toolkit gets the most attention, but driver support is a second factor in hardware end of life, and it explains Turing/Ampere's longer support horizon better than simply saying "CUDA 13 supports them".
- The driver branch. NVIDIA names R580 as the last driver branch for Maxwell, Pascal and Volta: it is a long-term branch with fixes until mid-2028, while the 590 and later branches no longer support these cards. Turing and newer architectures remain on the subsequent branches.
- Open GPU kernel modules. NVIDIA's open kernel modules work only from Turing upwards (they require the GSP processor, which Pascal and Volta do not have). In practice the P40/P100/V100 stay with the closed driver, while the T4/A2 can use the open one, which matters on newer Linux distributions.
This is no reason to panic (Pascal and Volta still receive security patches), but it shows fairly that the "future-proofing" of Turing/Ampere comes not from one but from two independent axes: the toolkit and the driver.
When it is worth paying extra for an A2 or T4
Not everyone should go for the cheapest card. There are three situations in which paying extra for Turing or Ampere pays off handsomely.
Compatibility with the newest toolchain. If you want to use the current stack, which targets CUDA 13, the A2/T4 have a compatibility advantage over Pascal and Volta: you will not be restricted to older CUDA builds as Pascal and Volta owners are, since CUDA 13 builds of PyTorch, llama.cpp or bitsandbytes cannot target their GPUs; individual libraries can still set a higher bar (native BF16 arithmetic and the main NVIDIA CUDA implementation of FlashAttention-2 require Ampere or newer, so the A2 meets that requirement and the T4 does not; BF16 emulation and a separate FlashAttention implementation for Turing do exist). But note: compatibility with CUDA 13 is not the same as "a good training card". Both have only 16 GB of VRAM, and the T4 additionally has compute capability 7.5 without BF16; this genuinely limits the scope of fine-tuning (model size, context length, training techniques). For more serious fine-tuning you will reach for a card with more memory anyway (we can source, for example, an A100 or L40 to order). The A2/T4 shine at light fine-tuning and inference on a modern stack rather than at heavy training.
A longer compatibility horizon. You buy once, put it in a server and do not want to keep replacing hardware. Turing and Ampere remain in official CUDA Toolkit 13.x support, so they give a longer horizon of compatibility with the new toolchain than Pascal and Volta, which are frozen on 12.x. This is not a guarantee that every future AI library will support them, but the starting point is clearly further ahead than with the older architectures.
Modern video and density. This is where the A2 shines brightest: a configurable 40–60 W power draw in a low-profile form factor (the T4 is low-profile too, at 70 W) means you can fit it where a full-size card will not go, and its power draw will not wreck the rack's power budget. Add hardware AV1 decoding: if you process modern video streams, that is a concrete, measurable advantage. The T4, in turn, offers hardware video encoding/decoding (NVENC/NVDEC) and has for years been a popular choice for transcoding and inference; it lacks only AV1 decoding.
And when a P40 or V100 is still enough
Now let us give the other side a fair hearing, because I do not want to sell you a more expensive card you do not need. If your use case is mainly LLM inference in GGUF format, that is, running ready-made, quantised models through llama.cpp and similar runtimes, Pascal and Volta can still be very useful on CUDA 12.x, especially when the priority is VRAM capacity relative to price. You do not need CUDA 13 for that.
What is more, this is where the older cards often win on what matters most for LLMs: the amount of VRAM for the money. The P40 24GB gives you 24 GB of memory and the V100 32GB as much as 32 GB, usually at a more favourable price per gigabyte than the A2 or T4, which have 16 GB each. If your bottleneck is whether a larger model will fit in memory at all, more inexpensive VRAM can matter more than the newest architecture. You will find more about why these cards still make sense for LLMs in our article on CUDA 13.
In short: the A2/T4 are the choice of a newer architecture and a longer compatibility horizon; the P40/V100 are the choice of more VRAM capacity for the money. Both make sense, for different needs.
Quick choice: application → recommendation
| Application | Recommendation |
|---|---|
| LLM / GGUF, maximum VRAM for the money | P40 24 GB |
| LLM / GGUF, the largest memory (32 GB) | V100 32 GB |
| Low power draw + modern toolchain | A2 16 GB |
| Transcoding / video encoding (NVENC) | A2 16 GB or T4 (to order): the same 7th-generation NVENC; the A2 also decodes AV1 |
| Hardware AV1 decoding | A2 16 GB |
| Newest CUDA / toolchain, peace of mind with drivers | A2 16 GB |
| Lowest cost of entry | P4 8 GB / P40 24 GB |
| Fine-tuning | A2 (with the 16 GB VRAM caveat); heavier training → A100/L40 to order |
The table is deliberately "not built around one card": for some applications the honest answer is a cheaper Pascal/Volta card, not the newest Ampere.
An honest disclaimer: all our cards are used, including the "future-proof" A2 and T4; CUDA support is not the same as a guarantee that a particular unit is in working order. Test every card on receipt (burn-in: several hours of load with temperature and stability monitoring) before you put it into production. Not sure whether the future-proof A2/T4 or the high-capacity P40/V100 will suit your application better? Write to us: we will help you choose honestly, without pushing a more expensive card on you.
Sources
- NVIDIA: CUDA Toolkit 13.0 Release Notes (removal of support for the Maxwell, Pascal and Volta architectures): https://docs.nvidia.com/cuda/archive/13.0.1/cuda-toolkit-release-notes/index.html
- NVIDIA: CUDA GPUs / compute capability (current list, A2 8.6, T4 7.5): https://developer.nvidia.com/cuda/gpus; older cards (P40 6.1, V100 7.0): https://developer.nvidia.com/cuda/gpus/legacy
- NVIDIA A2: datasheet (Ampere, 16 GB, 60 W, low-profile, AV1 decoding): https://www.nvidia.com/content/dam/en-zz/solutions/data-center/a2/pdf/a2-datasheet.pdf
- NVIDIA Tesla T4: datasheet (Turing, 16 GB, 70 W, NVENC): https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-t4/t4-tensor-core-datasheet-951643.pdf
- NVIDIA Tesla V100: datasheet (Volta, PCIe and SXM2 versions, 16/32 GB): https://images.nvidia.com/content/technologies/volta/pdf/tesla-volta-v100-datasheet-letter-fnl-web.pdf
- NVIDIA: Video Encode/Decode GPU Support Matrix (NVENC and NVDEC per board: every card here has NVENC; AV1 decoding only on the A2): https://developer.nvidia.com/video-encode-decode-support-matrix
- NVIDIA: Linux driver README, Open Linux Kernel Modules (Turing and later only, GSP firmware required): https://download.nvidia.com/XFree86/Linux-x86_64/560.35.03/README/kernel_open.html
- TechPowerUp: "NVIDIA's v580 Driver Branch Ends Support for Maxwell, Pascal, and Volta GPUs" (the 580 branch as the last with full support, security patches thereafter; Turing and newer unchanged): https://www.techpowerup.com/338497/nvidias-v580-driver-branch-ends-support-for-maxwell-pascal-and-volta-gpus
FAQ
- Will my P40/V100 stop working once CUDA 13 is out?
- No. It remains supported on the CUDA 12.x branch and continues to run inference, including GGUF models through llama.cpp. What changes is that the newest toolkit will no longer build code for these architectures, not that the card fails.
- Does CUDA 13 mean the T4 will be better than the V100?
- No. CUDA 13 gives the T4 an advantage in compatibility with the new toolchain, but it does not automatically mean higher performance. The V100 has 32 GB of VRAM (in the 32 GB version) and in many applications can be more practical if more memory is key. The choice therefore depends on whether your bigger constraint is software compatibility or memory capacity and performance.
- How do the A2 and T4 differ in practice?
- Both have 16 GB of VRAM, a low-profile single-slot form factor and the same 7th-generation NVENC encoder, and both remain supported by CUDA Toolkit 13.x. The A2 draws less power (a configurable 40–60 W vs 70 W), decodes AV1 in hardware and supports BF16; the T4 has much higher compute throughput and memory bandwidth. The choice depends on the power budget, the codecs you process and the speed you need.
- Is it worth paying extra for an A2/T4 for LLM inference?
- Not necessarily. If you mainly do inference in GGUF, the cheaper P40 24GB with more VRAM often gives better value for money. Paying extra for an A2/T4 makes sense when you care about fine-tuning on the newest stack, a long service life or newer codecs.
- Are the A2/T4 worth their price despite less VRAM?
- If you value future-proof support in CUDA Toolkit 13.x, low power draw and newer codecs, yes. If what matters to you above all is the amount of memory per model, older cards with 24/32 GB may be a better investment. It is a trade-off between a modern stack and raw capacity.
- What about fine-tuning on older cards?
- It can be done, but CUDA 12.x is the last toolkit series that can compile for these GPUs: PyTorch 2.14 with the CUDA 12.6 wheel is the last release with a prebuilt package for Pascal and Volta. These GPUs lack native BF16 arithmetic and support for the main NVIDIA CUDA implementation of FlashAttention-2; software BF16 emulation is a separate possibility. If you plan to use the newest training tools regularly, Turing/Ampere will spare you the compatibility struggle.