Unified Memory

Unified Memory is a design where the part of a computer that handles graphics and the part that handles general computing don't each get their own separate memory — instead, they both draw from a single shared pool of memory.

A regular computer is a bit like a kitchen where the chef (graphics processing) and the assistant (general computing) each have their own fridge. If one fridge is missing an ingredient the other has, you have to carry it over, and that takes time. Unified Memory is like merging those two fridges into one, so ingredients can be grabbed straight from the same space without any carrying back and forth.

This makes it easier to secure a large amount of memory all at once, which is especially useful when running today's large AI models on a personal computer. Devices like Macs or NVIDIA's DGX Spark use this kind of unified memory architecture, letting them run AI models tens or even hundreds of gigabytes in size without needing a separate, expensive server-grade graphics card.

An article describes Qwen's new model as running "on just 75GB of RAM or unified memory." It's easy to mistake unified memory for some special, separate type of memory, but it actually refers to a structure where graphics memory and the computer's overall system memory aren't kept separate — they're combined into one. That's why tables often clarify units by noting "total memory combining RAM and VRAM, or unified memory basis."