VRAM — Video Random Access Memory
Definition: High-bandwidth memory on a GPU that holds the model weights and activations during inference. VRAM — not raw compute — is usually the first ceiling you hit when you try to run a serious model locally.
Example
The 70B model was a non-starter at 24GB of VRAM; quantizing to 4-bit brought it down to 18 and the project lived.
When you'll hear it
VRAM shows up most often in AI strategy reviews, model evaluation discussions, and product roadmap meetings. When someone uses it, they're usually referring to video random access memory — and they expect the room to already know what that means.
FAQs
What does VRAM stand for?
VRAM stands for Video Random Access Memory.
What does VRAM mean in AI and machine-learning?
High-bandwidth memory on a GPU that holds the model weights and activations during inference. VRAM — not raw compute — is usually the first ceiling you hit when you try to run a serious model locally.
Where will I hear VRAM used at work?
VRAM comes up most often in AI strategy reviews, model evaluation discussions, and product roadmap meetings. It's used as shorthand for video random access memory, so people assume you already know the term.