GPUs for local AI
This index sits under /hardware, which answers "what can I run on this machine?". Each page below covers a single graphics card from the point of view of local model work: how much memory it carries, how fast that memory is, which apps run on it, and which models fit once you account for quantization.
These are card pages, not build pages. If you are choosing a whole machine rather than a single GPU, start with build guides. If you already have two cards in mind, go to comparisons.
Cards covered
- RTX 4090: what fits in 24 GB of GDDR6X, and the point where model size or context window pushes you past it.
- RTX 5090: what the larger 32 GB GDDR7 frame buffer changes about model choice, batch size and headroom.
- RTX 3090: why a previous generation 24 GB card still shows up in local AI builds, and what you give up against newer silicon.
- M3 Max: how unified memory changes the sizing question on Apple silicon, where the GPU and CPU share one pool.
- Radeon 7900 XTX: what 24 GB on AMD gets you, and how software support differs from the CUDA path.
How to read a GPU page
Every card page is organised in the same order, because that is the order the decisions actually happen in.
Memory capacity first. VRAM is the hard limit. A model either fits alongside its KV cache or it does not, and no amount of compute rescues a card that cannot hold the weights. Capacity sets the ceiling on parameters and on how long a conversation can run before you are trimming context.
Bandwidth second. Once a model fits, memory bandwidth is the main thing setting how fast generation feels. Single stream inference spends most of its time moving weights, not doing arithmetic, so a card with more bandwidth usually wins at the same capacity.
Runtime support third. A card is only as useful as the software that targets it. Each page says which apps run well on it, which formats such as GGUF are practical, and where support is partial. This matters more on AMD and Apple silicon than on NVIDIA, where most tooling is developed first.
Compute throughput comes last. It decides prompt processing, image and video generation speed, and fine-tuning time, but it rarely decides whether a project is possible at all.