A memory system where the CPU and GPU share one address space.
Unified memory is like one big kitchen counter. The cook and the blender grab from the same pile.
It helps local AI models move less data around. Speed still depends on memory bandwidth and smart scheduling.
GPU
Unified memory helps the GPU avoid some manual data copies.
VRAM
When VRAM is short, some data can use system memory.
Memory bandwidth
One shared address space does not mean the same speed.
Offloading
Offloading often uses this idea to move data between VRAM and system memory.