Dropped GPU support in CUDA and ROCm
Both GPU compute stacks retire old hardware. When they do, a card that worked on last year's release stops being a valid build target on this year's — and guides written before the cut keep saying it's supported. This page lists every retirement in this site'sCUDA and ROCm data, the last release that still targets each architecture, and the cards in the GPU collection each one affects: 8 NVIDIA and 6 AMD. Everything here is derived from the release data rather than written by hand, so it updates when a new release is added.
CUDA
4 NVIDIA architectures — Kepler, Maxwell, Pascal, and Volta — appear in earlier CUDA releases and not in CUDA 13.3, the newest tracked here. NVIDIA removes architectures only at major versions, so each retirement lines up with an x.0 release:
| Architecture | Compute capability | Newest CUDA that targets it | Removed in | Cards on this site |
|---|---|---|---|---|
| Kepler | 3.0 3.2 | CUDA 10.2 | CUDA 11.0 | — |
| Kepler | 3.5 3.7 | CUDA 11.8 | CUDA 12.0 | K80 |
| Maxwell | 5.0 5.2 5.3 | CUDA 12.9 | CUDA 13.0 | M40 24GB |
| Pascal | 6.0 6.1 6.2 | CUDA 12.9 | CUDA 13.0 | P100 SXM2 16GB, Tesla P40, TITAN X (Pascal), GTX 1080 Ti |
| Volta | 7.0 7.2 | CUDA 12.9 | CUDA 13.0 | TITAN V, V100 SXM2 32GB |
Kepler went in two steps: compute capability 3.0/3.2 was removed in CUDA 11.0, and 3.5/3.7 — the K80 among them — held on through the 11.x series until 12.0.
The most recent cut: Maxwell, Pascal, and Volta in CUDA 13.0
CUDA 13.0 removed Maxwell, Pascal, and Volta in a single release — every compute capability below 7.5. That is the retirement most likely to affect a card you own or are thinking of buying: it covers M40 24GB, P100 SXM2 16GB, Tesla P40, TITAN X (Pascal), GTX 1080 Ti, TITAN V, and V100 SXM2 32GB, several of which are still bought second-hand for their VRAM.
What "removed" means in practice: CUDA 12.9 and earlier 12.x toolkits can still build for these cards, and NVIDIA states that building for them with 12.x remains supported. CUDA 13.0 and later cannot target them at all. So the working setup for one of these cards is a 12.x toolkit, plus framework builds made against 12.x.
Two caveats. First, the toolkit and the driver are separate: this page tracks which toolkit can compile for an architecture, not which driver branch still runs it. Second,prebuilt framework binaries make their own choices. A PyTorch wheel or a llama.cpp release is compiled for a specific list of architectures, and a project can leave old ones out of a build even while the toolkit it uses still supports them. If a card is on this table, check the architecture list of the exact build you install.
ROCm
ROCm grants support per gfx target rather than per architecture, and the same question asked of the ROCm releases tracked here (4.5.0 through 10.0.0) returns one target:
| gfx target | Supported from | Newest ROCm that lists it | Removed in | Cards on this site |
|---|---|---|---|---|
| gfx906 | ROCm 4.5.0 (or earlier) | ROCm 6.3.3 (deprecated from 6.0.0) | ROCm 6.4.0 | Radeon Instinct MI50, Radeon Instinct MI60, Radeon VII |
Older Instinct cards were retired before ROCm 4.5.0, where this site's ROCm data begins, so no tracked release lists them: Radeon Instinct MI25 (gfx900), Radeon Instinct MI6 (gfx803), and Radeon Instinct MI8 (gfx803). The exact release that dropped each one is outside that range and isn't asserted here.
The pinning advice is the same as for CUDA: a card on this table needs a ROCm release from before its cut, plus framework builds made against that release. The difference is that the whole older ROCm line is end-of-life, so pinning means no further fixes. For llama.cpp-class inference there is also a way round the question entirely — a Vulkan build, which needs a working Vulkan driver rather than a place in AMD's support matrix. SeeROCm on consumer GPUs for what that route covers and what it doesn't.
"Deprecated" in the table means AMD still shipped support for the target in those releases but had announced its removal, so they are the newest ones to pin. AMD's current pages for the first release without a target can also differ from what was published at launch: the 6.4.0 requirements page now lists gfx906 as unsupported, but an archived copy from May 2025 showed the Radeon VII and Radeon PRO VII as deprecated.
Outside official releases, AMD's TheRock build system still publishes nightly ROCm packages for some retired targets, gfx906 and gfx900 among them. Its support table lists them at the lowest tier, "Build Passing", which TheRock defines as a package being produced, not as working on the hardware. Some owners report good results with them; treat them as unverified.
The takeaway
"Supported" is a property of a release, not of a card. Before buying an older GPU or following an older guide, check the card against the release you'll actually install: eachGPU page lists every CUDA or ROCm version that covers it.