What you have in a graphics card under Linux: running AI yourself, and gaming
There are two reasons to put a graphics card in your machine that have nothing to do with office work. The first is that you want to run a language model or image generator on your own computer, instead of sending your text to a service on the other side of the ocean. The second is that you want to play. Under Linux, both are possible nowadays, and that was a very different story five years ago.
This piece is about the question you have to answer before buying such a machine: how much memory should be on that card?
The memory on the card is the measure, not the speed
When it comes to a graphics card, most shops look at frames per second. For AI, that is the wrong number. What counts is the memory on the card itself, the video memory, because a model has to fit entirely in it. If it doesn't fit, it falls back to your regular working memory and to the processor, and then it becomes ten to twenty times slower. Not broken, but unusable.
The rule of thumb is simple. A model is stored with a certain precision, and the common form to run at home is four bits per weight. Then roughly: half a gigabyte per billion parameters. A model of seven billion parameters therefore costs about four gigabytes, and on top of that comes some extra for the conversation you have with it. The longer the conversation, the more.
What fits where
| Video memory | Language models | Image |
|---|---|---|
| 6 GB | models up to about 8 billion parameters, with short conversations | the older image models, fine results, no large formats |
| 8 GB | the same models, but roomier: longer conversations without hassle | also the larger image models, provided you have patience |
| 12 GB | models around 13 billion parameters | comfortable, also larger formats |
| 16 GB | models up to about 14 billion comfortably, or a smaller model more accurately | the heavier image models, and several at once |
| 24 GB and more | the class of 30 billion parameters | almost everything that is feasible at home |
There is a jump in that table that you need to see. Between 8 and 16 gigabytes is not "a bit more room" but a different category of models. Anyone who today hesitates between those two is actually buying the answer to the question of whether he will still be satisfied in a year.
How to get it working
The simplest way is Ollama. That is one program that downloads the models, puts them in memory, and has a conversation with them, without you having to build anything.
curl -fsSL https://ollama.com/install.sh | sh ollama run llama3.2
After that, you talk to the model in your terminal. If you want a window instead of a terminal, there are various shells that talk to the same models. For image, ComfyUI is the usual choice; it requires a bit more setup but lets you see exactly what happens.
On an Nvidia card, under Linux you need the closed driver from the manufacturer. On Ubuntu and Debian, that installs with a few lines and then just works: for AI, and just as well for gaming. That last part was the weak point for years, and it no longer is. AMD cards use an open driver that is already in the kernel, which is nicer if you want to keep everything open; for AI, the range of tools for Nvidia is still larger. For both, the following applies: it works, and the choice is about what you consider important.
And gaming then?
That is the part that has quietly turned around in recent years. Valve, the company behind Steam, has spent years building Proton: a layer that lets games made for Windows run under Linux. That started as a novelty and has grown into something their own handheld, the Steam Deck, relies on completely. That machine runs Linux, and precisely because of that, the number of games that work has grown enormously: a manufacturer who wants to sell his game on the Deck makes sure it works under Proton.
In practice, you install Steam on your Linux machine, check in the settings that Proton may be used for all games, and most things just start. On the site ProtonDB, players track per game how well it goes.
One thing you should know, and we are honest about it: games with anti-cheat software that reaches deep into the system often do not work. That mainly affects the big online shooters. If you play those, first check your list before you switch. The rest of your library works, and often just as fast as under Windows.
So what do you buy
If you mainly want to try out what is possible at home, and you don't mind staying with the smaller models, then six or eight gigabytes is enough. If you want to run a model that truly thinks along with your work, or you want to enjoy it for longer, then sixteen gigabytes is the difference between "it fits" and "it doesn't fit".
And do you already have a computer that is otherwise fine? Then you don't need to buy a new one. Just put a card in it and you are done: the same machine, new tools. That is, by the way, exactly why we offer it that way, and not as a package where you are not allowed to change anything.
Whatever you choose: it runs at your home. Your conversations go nowhere, there is no subscription that gets more expensive, and there is no service that one day decides that your question is not allowed. That is the same reason why we sell servers for which you have the key.