Skip to content

Unified memory

Short answer

Unified memory is a design where the CPU and GPU share one pool of RAM instead of each having their own, as on Apple Silicon and in integrated graphics.

For local models this is often an advantage: a machine with 32 GB of unified memory can give far more of it to a model than a desktop GPU with 8 GB of dedicated VRAM, with no transfer across a bus.

The bandwidth is usually lower than a discrete card's, so the pattern is that unified-memory machines can run bigger models than their price suggests, somewhat more slowly than their specification suggests.

Related terms

Other ways people ask this

  • what is unified memory

Related