Definition
Plain language
The dedicated fast memory built into a graphics chip, separate from the computer's main memory.
As stated in the literature
Video RAM — the high-bandwidth on-card memory of a GPU that holds model weights, activations, and the KV cache during training and inference; its capacity is the binding constraint that forces workarounds like CPU offloading and quantization.
Why it matters: Its limited capacity is often the hard wall that dictates how big a model you can run and forces tricks like offloading or shrinking the model to fit.
For example, a large model may simply refuse to load if it needs more VRAM than the graphics card physically has.
Heard on the show
“You keep the optimizer's bookkeeping in regular system RAM instead of GPU memory, because there's just not enough VRAM.”Episode 009 — How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers