VRAM
Also called: video memory, GPU memory, graphics memory
The memory built into a graphics card, where the GPU stores the data it processes. In AI, its capacity sets the size of the models you can run.
VRAM, or video memory, is the memory installed directly on the graphics card, around the graphics processor. It is separate from the computer's main RAM and much faster. It is measured in gigabytes: 8 GB on an entry-level card, several dozen on professional cards, and more than 100 GB on some data center GPUs.
For artificial intelligence, VRAM is often the number one criterion. A language model has to fit in memory, along with its parameters, its context cache and the data being processed. As a rough guide, each billion parameters takes up about 2 GB at 16-bit precision, and two to four times less once quantized. If there is not enough VRAM, the model will not load or becomes very slow. For mining, the amount of VRAM also matters for some memory-hungry algorithms.
On a product page, look at the VRAM capacity, its type, for example GDDR7 or HBM, and its bandwidth. To run a model locally, start from the size of the model you are aiming for and add a margin for the context window. With several GPUs, memory can be split across the cards, but the link between them then becomes important.
Our AI GPU collection and our dedicated page help you choose the right amount of VRAM for your models.