VRAM

VRAM is the memory built onto a graphics card, physically separate from system RAM and wired to the GPU die over a wide local bus. It holds whatever the GPU is working on: textures and frame buffers for graphics, model weights and activations for AI. Running out of it is a hard stop rather than a slowdown, since there is no transparent overflow into system memory.

Consumer and workstation cards use GDDR6, GDDR6X, or GDDR7, reaching 300 GB/s to roughly 1.8 TB/s depending on bus width and speed grade. Data center accelerators use HBM, which stacks memory dies on the same package as the GPU for both more capacity and more bandwidth. An H100 with 80 GB of HBM3 moves about 3.3 TB/s. Capacity runs from 8 GB on entry gaming cards to 96 GB on the top workstation part, and 80 GB or more on data center accelerators. A 7 billion parameter model at 16-bit precision needs about 14 GB for weights alone, before context or any batch above one.

Size to the largest model or scene that must be resident at once, with headroom. For AI work, check a model's quantized footprint against card capacity before assuming a GPU can run it.

Sources

Last verified August 29, 2026.