Skip to content
Fathom Layer
GlossaryGlossary7 MIN READ

What is unified memory, and why does it matter for AI?

Compact board with processor and adjacent memory packages under a removed heatsink
AI-generated editorial illustration · Conceptual scene

Unified memory is a design in which the CPU and GPU share one pool of system memory instead of each having its own, so the GPU can work on data without it first being copied into separate video memory. Apple introduced it to the Mac with M1, describing it as high-bandwidth, low-latency memory in a single pool that lets all the technologies on the chip "access the same data without copying it between multiple pools of memory". AMD uses the same term for its Ryzen AI Max processors. For local AI, the practical effect is capacity: a GPU on unified memory can address far more memory than a consumer graphics card carries.

How it works

In a PC with a discrete graphics card, the CPU uses system RAM and the GPU uses its own VRAM. Data the GPU needs has to be moved into VRAM, and the model has to fit there to run entirely on the GPU.

With unified memory there is one pool. Apple's Metal documentation puts it plainly: "Apple GPUs have a unified memory model in which the CPU and the GPU share system memory." Access still has rules. Metal lets developers mark memory as shared by both processors or private to the GPU, and Apple exposes an estimate of how much memory the GPU can allocate "without affecting its runtime performance" (recommendedMaxWorkingSetSize). Apple does not publish what that limit is for a given Mac, so the GPU cannot count on every gigabyte of the pool.

AMD's version works through a firmware setting. AMD describes Variable Graphics Memory as a BIOS-level feature that reallocates a share of system RAM to the integrated graphics, made possible by the unified memory architecture of its Ryzen AI processors. On the Ryzen AI Max+ 395, AMD says up to 96 GB of a 128 GB configuration can be converted to VRAM this way.

A note on the name: NVIDIA's CUDA also has a feature called Unified Memory, which NVIDIA defines as "a single memory address space accessible from any processor in a system". That is a programming model that can span separate CPU and GPU memories, not the same thing as a chip whose CPU and GPU physically share one pool.

Capacity and bandwidth, from the official pages

Memory bandwidth matters because generating each token requires reading the model's weights from memory; see tokens per second. The figures below are the manufacturers' published numbers, checked on 29 September 2026; this index did not measure them.

System Maximum memory Memory bandwidth (published)
Apple Mac Studio, M5 Ultra 512 GB unified 1.2 TB/s
Apple Mac Studio, M5 Max (40-core GPU) 128 GB unified 614 GB/s
AMD Ryzen AI Halo developer platform (Ryzen AI Max+ 395) 128 GB LPDDR5x (up to 96 GB to GPU) 256 GB/s
NVIDIA DGX Spark 128 GB coherent unified system memory 273 GB/s
NVIDIA GeForce RTX 5090 (discrete, for comparison) 32 GB GDDR7 1,792 GB/sec

Apple's MacBook Pro specifications list bandwidth from 153 GB/s to 614 GB/s depending on the chip, so "a Mac with unified memory" covers a wide range. AMD has also announced the Ryzen AI Max PRO 400 Series, which it says offers up to 192 GB, with up to 160 GB dedicated to the GPU.

What it changes for AI workloads

  • Larger models on one device. A model's weights take parameters × bytes per parameter; at 16-bit that is 2 bytes per parameter, so a 70-billion-parameter model needs about 140 GB for weights alone (our arithmetic). That exceeds any single consumer card in the table but fits in the larger unified-memory configurations. NVIDIA says DGX Spark's 128 GB lets developers work with models up to 200 billion parameters; AMD says its 128 GB Ryzen AI Max+ systems can load models up to 235 billion parameters with its recommended quantization. Both are the vendors' own claims about their own products.
  • No split between CPU and GPU memory. On a discrete card, a model that overflows VRAM must be split, with part running from system memory. Unified memory avoids that particular split, within the limit the operating system or firmware allows the GPU.

The trade-off

Capacity and bandwidth pull in different directions. On the published numbers above, the RTX 5090 has more bandwidth than any unified-memory system listed, but 32 GB of memory. The M5 Ultra has 16 times the memory of the RTX 5090 and about two-thirds of its bandwidth (our arithmetic from the two official figures). The AMD and NVIDIA desktop systems have 128 GB with a fraction of that bandwidth.

Inference from those numbers, not a measurement: for a model that fits comfortably on the discrete card, the card's higher bandwidth should give faster generation; for a model that does not fit, unified memory can run it on one device where the card cannot. Real speed also depends on the software stack and on compute, which the bandwidth figure does not capture.

When it fits, and when it does not

Good fit: running large quantized or 16-bit models on a single machine; development work where capacity matters more than peak speed.

Weaker fit: workloads where a model fits in a discrete card's VRAM and speed is the priority; software that depends on a specific GPU platform, such as tools built only for CUDA. Check that your inference engine supports the hardware: llama.cpp, for instance, calls Apple silicon "a first-class citizen", while other engines target NVIDIA GPUs first.

In the index

Sources

END OF ANALYSIS

In the Fathom index

Systems in the index that list unified memory. Matched on each item's own record and sorted A–Z, not by merit.

Showing 4 of 9.Browse Local AI hardware →

Related Intelligence