How we calculate AI memory

AI memory is the single most useful number on this site, and it is the one that hardware listings get wrong most often. A machine advertised with 128GB of memory does not give a local model 128GB to work with. How much it really gives depends entirely on the memory architecture, and the gap between the advertised number and the usable number can be more than half.

Discrete GPUs

On a desktop or laptop with a discrete NVIDIA card, AI memory is the VRAM on the card. A machine with an RTX 5090 and 64GB of system RAM has 32GB of AI memory, not 96GB. System RAM still matters, because a model can be split so that some layers run on the processor, but that path is slow enough that we list it separately as CPU offload capacity rather than folding it into the headline figure.

Unified memory on Ryzen AI Max

On Strix Halo systems the graphics processor draws from the same pool as the processor. The Windows Variable Graphics Memory default hands the graphics side roughly three quarters of the pool, so a 128GB machine reports about 96GB, a 96GB machine about 72GB, a 64GB machine about 48GB and a 32GB machine about 24GB. We use that three quarter rule. On Linux the allocation can be pushed higher, to around 110GB on a 128GB machine, and the system page says so where it applies.

Unified memory on Apple silicon

Apple systems share one pool between processor and graphics, and macOS applies a wired memory limit that defaults to about three quarters of the total. A 128GB Mac Studio therefore offers roughly 96GB to a model out of the box. That ceiling can be raised with a system setting, and people running large models routinely do, but we score the default because that is what a buyer gets without intervention.

Coherent memory on NVIDIA GB10

GB10 systems such as DGX Spark and its partner machines use a coherent pool where the graphics processor addresses almost all of the memory directly. The practical reservation is the operating system itself, around 8GB, so a 128GB machine offers about 120GB and a 64GB machine about 56GB. This is why these systems hold the top of our memory rankings despite modest bandwidth.

Shared integrated graphics

Intel Core Ultra, Snapdragon X and the non-Max Ryzen AI 300 parts share system memory with integrated graphics, and the usable share is roughly half. These are genuinely useful machines for the AI features built into the operating system and for small models, and we classify them as AI Productivity PCs so that the distinction is visible rather than buried.

Why we enforce a floor

A system with less than 16GB of usable AI memory cannot be classified on this site as a Local AI PC, an Agent Computer or an AI Workstation, whatever its marketing says. Below that line the useful local model options run out quickly, and labelling such a machine as an AI system for local work would mislead the person reading the page.