Unified memory
Unified memory is a single pool of RAM that the CPU and GPU both address directly, with no separate graphics memory and no copying between the two. Apple silicon chips are the best known example, and the term comes up constantly in local AI discussion, so it earns its place in the glossary.
On a conventional desktop the GPU has its own VRAM and the CPU has system RAM. Moving data between them costs time. On a unified memory system there is one pool, and a model loaded there is visible to both processors as it sits.
Why it matters for local AI
The practical consequence is capacity. A discrete card ships with a fixed amount of graphics memory that you cannot expand, and that amount is usually much smaller than the system RAM in the same machine. A unified memory machine can devote a large share of a large pool to a model, so it can hold models that simply will not fit on a consumer GPU.
The trade is speed and ecosystem:
- Memory bandwidth on unified memory parts is generally lower than on a high end discrete GPU, so generation is slower even when the model fits comfortably.
- Bandwidth varies a lot across a single chip family, so the tier of chip matters as much as the amount of memory.
- Some tooling is built GPU first, and support for unified memory backends can lag.
It is a good fit for large models at modest speed, and a poor fit for workloads that need raw throughput, such as batch inference or heavy image and video work.
Rule of thumb
Buy unified memory for capacity, buy a discrete GPU for speed. If your goal is running the largest model you can at a readable pace, a big unified pool wins. If your goal is fast generation or training, it does not.
Not all of the pool is yours. The operating system and your open applications hold a meaningful share, so plan around usable memory rather than the number on the box.