VRAM
High-bandwidth memory on discrete graphics cards
Hardware
In plain English
The ultra-fast backpack memory attached directly to your graphics card.
Technically
Dedicated GPU memory (GDDR6X/GDDR7 or HBM3e) offering 500 to 3000+ GB/s bandwidth. Model weights + KV cache must fit within VRAM to avoid the massive 30x bandwidth penalty of offloading layers over PCIe.