Graphics card prices for theLocal AI

Which graphics card should you buy to run an LLM locally?

Look at video memory (VRAM) first: the model must fit there entirely, or it spills into system memory and speed collapses. In Q4, 12 GB is enough for models up to about 14 billion parameters, 16 GB up to about 24 billion, 24 to 32 GB for models with 30 to 35 billion, and 48 GB up to about 70 billion. With equal memory, bandwidth is the deciding factor: it determines generation speed.

NVIDIA, AMD, or Intel for local AI?

NVIDIA remains the simplest choice: CUDA is supported by all the tools (Ollama, LM Studio, llama.cpp, vLLM). Radeon RX 9000 and AI PRO R9700 are supported by Ollama (ROCm 7), LM Studio and llama.cpp, sometimes with a bit more configuration. Intel Arc cards mainly use Vulkan or SYCL (llama.cpp, LM Studio) and OpenVINO: they often offer the cheapest VRAM, but the least mature ecosystem.

Is the RTX 5090 worth its price for AI?

This is the fastest consumer card (1,792 GB/s) with 32 GB, but it cost €6,699.95 in the latest check, compared with €2,349 at launch. For 32 GB, a Radeon AI PRO R9700 (€2,299.95) or an Intel Arc Pro B70 (€2,299.95) costs much less, at about 2.8 times lower speed. For very large models, a mini-PC with 128 GB of unified memory remains the alternative.

Where do these prices come from, and when are they updated?

Automatically collected every Monday and Thursday from our partner Materiel.net's product feed. For each chip, we select the least expensive card in stock or shipping within 7 days, across all brands. Amazon prices are not archived: Amazon only allows recent prices to be displayed, so they appear live on the buttons. The links are affiliate links: a purchase earns us a commission at no extra cost to you and does not influence the ranking, which is calculated automatically.

How is speed estimated?

To generate each token, the card rereads the entire model in memory: maximum speed is roughly bandwidth divided by model size. We use 70% of that figure for a 14-billion-parameter model in Q4 (about 9 GB). Actual measurements vary by software and drivers: this is an order-of-magnitude estimate for comparing cards, not a promise.