Compare Graphics Cards for AI
Put a workstation card head-to-head with any other GPU on what matters for running models on your own machine: memory, the largest model each one holds, generation speed, render and compute benchmarks, and what each GB of memory costs today.
Prices update daily from live retailer stock
Cards marked out of stock have no live offer right now. You can still compare them on specs and benchmarks.
How to compare two graphics cards for AI
Start with memory. A model either fits in a card’s memory or it does not run on that card at all, so the first thing each comparison shows is the largest model each card holds, at the usual compression and at two higher-quality levels, and how many of each card a 70B model takes.
Then speed. Every word a model generates reads the whole model from memory, so generation speed follows memory bandwidth far more closely than it follows raw compute. Two cards with the same bandwidth generate at about the same pace however different their price, which is why a comparison can show a much cheaper card keeping up.
Then cost per GB, from the lowest price we are able to quote today, and the software each card runs: most AI tools support CUDA first, so a card on another stack can be the better buy on paper and still the wrong one for the work. Render, compute and 3D benchmark scores and the full specifications follow, row by row.
Not sure which cards to shortlist? The ranking of graphics cards for AI sorts every card by what it can run and what it costs, including setups that pool two or more cards. Comparing for games instead? Use the graphics card comparison, which is built around frame rates.