Unified Memory Architecture

A computer has working memory. This is the fast intermediate storage where all the data currently being processed resides. Classic PCs have two separate areas for this: one for the main processor, which handles general computing tasks, and a second one on the graphics card. In a Unified Memory Architecture, this separation does not exist. All computing units access a single, shared memory. The principle became well known through Apple's M chips, but other manufacturers now build this way too.

Why copying costs so much time

In the classic setup, data must be shuttled back and forth between both memories. An image that the graphics unit is supposed to process is first copied from the main memory into the graphics memory. Afterward, the result travels back again. These copying operations run over a connecting link that is noticeably slower than the memory itself.

A comparison makes this vivid. Two cooks work in separate kitchens and must carry every ingredient through a long corridor. Even if both chop quickly, the corridor determines the pace. Shared memory is like one large kitchen with a single work surface for both. The ingredients simply lie where they are needed.

For AI models, the sheer amount also matters. A large language model needs many gigabytes of space just to fit into memory at all. A normal graphics card often has 8 or 16 gigabytes permanently built in, and no more is possible. A notebook with unified memory, on the other hand, can have 64 or 128 gigabytes, and the graphics unit is allowed to use a large portion of that. That’s why some AI models run on a laptop that won’t even start on a more expensive desktop PC.

One address space for all computing units

Technically, the main processor, the graphics unit, and often a special AI computing unit sit on a single chip. The memory is connected right next to it, with very wide data lines. All units address the same memory addresses. If the main processor stores something at address X, the graphics unit finds it at exactly the same address.

Instead of copying data, a program now only passes on a reference. It’s as if, instead of sending the document itself, one merely said which shelf it’s on. This avoidance of copies is called zero copy and is the actual core of the concept.

One disadvantage follows directly from this: the memory is permanently soldered and cannot be upgraded. Anyone who needs more later has to replace the entire device. Furthermore, all units share the same bandwidth, meaning the same amount of data per second. If the graphics unit and processor both read heavily at the same time, they slow each other down. Another common misconception exists: unified memory does not automatically make a chip faster than a large graphics card. It only removes a particular bottleneck.

From Apple Silicon notebooks to the data center

The term is most commonly encountered in connection with Apple. Since 2020, MacBook spec sheets have advertised Unified Memory, and reviews regularly mention the advantage for video editing and local AI models. Gaming consoles like the PlayStation 5 have also long worked with shared memory. AMD and Intel are building the principle into their chips for notebooks.

In finance and tech news, the topic usually comes up in connection with AI hardware. Nvidia sells systems in which the processor and graphics chip see a shared memory area, because otherwise the models would no longer fit. How much memory a chip can serve has become an important selling point there.

For a buyer, this means in practice: with a device that has shared memory, it’s better to plan generously when purchasing. The stated figure applies to everything combined, not just to programs. Anyone who wants to experiment with AI locally will otherwise quickly hit a limit that can no longer be shifted.