Definition
Unified memory is Unified memory is a computer design where the processor and graphics chip share one pool of memory, which lets a laptop or desktop load AI models far larger than a typical graphics card holds.
Unified memory is a way of building a computer so the CPU, the GPU and other processors all draw from the same pool of memory. Apple's Mac chips are the best-known example, though other chipmakers sell similar designs.
In many desktop computers, the CPU uses system memory and the graphics card has its own separate VRAM. If data needs to move between them, it has to be copied across a connection, which takes time and space. A graphics card might have 8 to 24 gigabytes of its own memory, and that number caps how big a model you can run on it.
With unified memory, there is one pool. The GPU can use as much of it as the system allows, and nothing needs to be copied between separate pools. If a machine has 128 gigabytes in total, a large share of that can hold a model.
Running a language model locally is mostly a memory problem. The model's weights must fit in memory, with room left for the KV cache, which grows as your prompt and conversation get longer. Quantization shrinks the weights, but a model with many billions of parameters still needs a lot of space.
A rough rule: at 4-bit precision, each parameter takes about half a byte. So a model with 100 billion parameters needs around 50 gigabytes for weights alone, before the cache and everything else your computer is doing.
Unified memory lets a laptop hold models that would otherwise need an expensive multi-GPU server. That is the main reason people use these machines for local AI.
Capacity is not speed. Unified memory usually has lower bandwidth than the memory on a top data center GPU, so a model may fit but generate text more slowly. Bandwidth, how fast data moves between memory and the chip, often decides how many tokens per second you get.
You also cannot upgrade it later. On most machines with unified memory, the amount is fixed when you buy, so you have to guess your needs ahead of time.
Look at total memory first, then bandwidth, then independent speed tests on the specific model you plan to run. Vendor claims about "large models on device" tell you what fits, not how fast it runs.
For related ideas, see inference, the act of running a trained model, and high-bandwidth memory, the fast stacked memory used in data center AI chips.
No. VRAM is memory dedicated to a graphics card. Unified memory is one shared pool used by both the CPU and the GPU.
Roughly half a byte per parameter at 4-bit precision, plus room for the context cache and your other apps. A 70-billion-parameter model needs about 35 gigabytes for weights alone.
Microsoft made its agent containment layer generally available on Windows 11 and opened preorders for Nvidia RTX Spark PCs, from $2,599.99 for the Surface Laptop Ultra. The laptop gets the headlines, but the fence around your agents is the bigger change.
Apple's new MacBook Pro with M5 Pro or M5 Max adds a touch OLED screen and up to 128GB of memory, from $2,899, shipping Nov. 6. Apple says M5 Max runs models with hundreds of billions of parameters locally but gives no speeds, so wait for real numbers.
Meta released Muse Spark Open, an 816B-parameter open-weight model it says runs on two M5 Ultra Mac Studios. The size is real. The scorecard and the full license terms are still missing.