Glossary › AI Computing › VRAM
What is VRAM?
VRAM (Video RAM) is the dedicated high-speed memory on your graphics card. For AI, it’s the single spec that decides which models you can run at all.
Why VRAM is king for AI
An AI model must fit entirely in VRAM to run at full speed. A 7B-parameter LLM needs ~5–6GB; a 13B model needs ~10GB; image generators like SDXL want 8–12GB. Run out of VRAM and things slow to a crawl or fail entirely.
VRAM tiers in our builds
8GB (RTX 3070 Ti, $950): 7B–8B LLMs and Stable Diffusion. 12GB (RTX 4070, $1,500, or RTX 3080 Ti / 5070): 13B models comfortably. 16GB (RTX 5070 Ti / 5080 class): larger quantised models up to ~30B and serious fine-tuning. For gaming, 8GB handles 1080p/1440p; 12GB+ helps at 4K.
VRAM on every RTX card
VRAM doesn’t always follow performance — an RTX 3060 has 12GB while the faster RTX 4060 has only 8GB. See the VRAM of every card from the RTX 2060 to the 5080, side by side with its raw power score, in our RTX GPU comparison.
Common questions
RAM is the main memory the CPU uses for Windows and programs. VRAM is separate memory on the graphics card, used for game textures and AI models. They are not interchangeable, and you cannot add VRAM later because it is fixed to the card.
Probably a few more years at 1080p and 1440p, though nobody can promise that. 12GB is comfortable today at those resolutions and runs short first at 4K with maxed textures or with path tracing. For local AI, 12GB runs mid-sized models but not the largest ones.
For gaming, yes: no current game needs it. For running large AI models on your own PC, training, or heavy 3D and video work, 32GB is useful and sometimes necessary.
More questions? See all common PC questions.