VRAM

No jargon, no math

An AI model is a huge table of numbers (its parameters), and running it means reading that whole table for every answer. VRAM is the memory built to sit millimetres from the compute cores, on a very wide connection, so thousands of numbers flow per clock tick. What goes in: the model file plus the question. What happens: every layer of the model multiplies its slice of the table against your input, storing intermediate results in the same memory. What comes out: the answer, and a memory footprint you can measure. Where it breaks: capacity is a wall, not a slope. A model whose weights plus working space exceed your VRAM does not run slower, it refuses to load. That is why a 70-billion-parameter model needs roughly 140 GB at full precision, and why teams shrink models with quantization to fit smaller cards.