Dedicated GPU memory, determining what size AI model can be loaded and run.
VRAM is memory attached directly to the GPU. This is a hard limit when running an AI model: if the model and intermediate data do not fit in VRAM, it cannot run, no matter how powerful the GPU is.
Rule of thumb for inference: a 4-bit quantized model requires roughly half the number of parameters in GB. Training costs several times more than inference because gradients and optimizer state must be retained.
When comparing GPU plans, compare VRAM first, then computational performance — insufficient VRAM means no execution, while insufficient speed only means slower execution.
