Unified Memory vs VRAM for AI
Updated
These two architectures are routinely compared on capacity alone, which hides the thing that actually separates them. A 128GB unified machine and a 32GB graphics card are not a bigger and a smaller version of the same idea.
What is the real difference?
A discrete graphics card has its own memory attached directly to the processor on a very wide bus. That delivers enormous bandwidth, currently up to about 1,792 GB/s, but the capacity is fixed at manufacture and cannot be extended.
Unified memory puts one pool behind both the processor and the graphics processor. Capacity can be far larger, up to 512GB on the biggest machines, but it is reached over a narrower path, typically between 256 and 1,200 GB/s depending on the design.
Which one runs a bigger model?
Unified memory, and it is not close. The largest consumer graphics card holds 32GB. A unified machine reaches 96GB, 120GB or 384GB of usable AI memory. If your target model does not fit a card, no amount of card bandwidth helps, because the model simply will not load.
Which one is faster?
A discrete card, on anything that fits in its memory, usually by a wide margin. Generation speed tracks memory bandwidth closely, so a card reading memory five times faster produces text roughly in proportion. For image and video generation, which are compute-bound rather than memory-bound, the gap is larger still.
So which should you buy?
Work out the largest model you intend to run. If it fits comfortably inside a card, buy the card and enjoy the speed. If it does not, buy unified memory and accept a slower stream of text, because the alternative is not running the model at all.
What happens when a model does not fit?
It still runs, which is the source of most confusion. Runtimes will place as many layers as fit on the graphics processor and run the rest on the processor using system memory. The model loads and produces text, so nothing appears broken.
What changes is speed, and the drop is severe rather than gradual. Processor memory bandwidth is a fraction of graphics memory bandwidth, so the offloaded layers dominate the total time. A model that is 20 percent offloaded can easily run several times slower than one that fits entirely. This is why we treat a fit as comfortable only with headroom, and label anything tighter as limited.
Does the operating system change the usable figure?
Substantially, and differently on each platform. Windows hands a Ryzen AI Max graphics processor about 75 percent of system memory by default, and that split is adjustable in firmware on some machines. macOS applies a wired memory limit that defaults to around 75 percent and can be raised with a system setting. A GB10 machine reserves only the operating system itself, around 8GB of 128GB.
We score the default on every platform, because that is what a buyer gets without intervention. Where raising the limit is practical, the system page says so.
Which architecture ages better?
Unified memory, on current evidence. Model sizes have grown faster than consumer graphics memory, and a card's capacity is fixed at manufacture. A machine bought today with 96GB of usable memory will still hold next year's mid-sized models; a 16GB card already cannot hold this year's.
The counter-argument is real though: bandwidth has improved faster on the discrete side, and software support for CUDA remains ahead. If your models are small and your tooling is NVIDIA-shaped, a card remains the better experience.
A worked comparison
Put a 32GB discrete card next to a 128GB unified machine. The card reads memory at up to 1,792 GB/s and holds a 32B model at Q4 with room to spare, generating quickly. The unified machine reads at 256 GB/s and holds a 70B model comfortably, generating at perhaps a fifth of the speed.
On a 32B model the card wins clearly and the comparison is about speed. On a 70B model the card cannot compete, because it cannot load the model at all. Both statements are true at once, which is why a single ranking of these two architectures is misleading and why this site ranks them per workload.
What about professional cards?
An RTX PRO card with 48GB or 96GB of ECC memory is the one option that gets both: large capacity and full discrete bandwidth. It holds a 70B model entirely in fast memory and runs it faster than any unified machine.
The price reflects that, and it is the reason the professional tier exists. If both capacity and speed are genuine requirements rather than preferences, this is the answer and the other two architectures are compromises.
How to decide in one step
Name the largest model you intend to run in the next two years and apply the memory formula. If the answer fits a card you can afford, buy the card. If it does not, buy unified memory. Every other consideration on this page is secondary to that single test.
How we calculate the figures on this page
The comparison on this page rests on two measured numbers for every machine: usable AI memory and memory bandwidth. Usable AI memory is what a model can actually occupy after the platform takes its share, and the rule differs by architecture rather than by brand. A discrete card gives its full VRAM and nothing more, because system RAM is reachable only through a slow path we record separately as CPU offload capacity. A Ryzen AI Max machine under Windows hands the graphics processor about three quarters of system memory by default. Apple silicon applies a wired limit at roughly the same proportion, raisable through a system setting that we do not assume anyone has changed. An NVIDIA GB10 machine reserves only the operating system, around 8GB. A shared-graphics laptop gives about half. Memory bandwidth is taken from the vendor specification and stored per configuration where it varies within one chip name, which it does on Apple's M5 Max: 460 GB/s with the 32-core graphics processor and 614 GB/s with the 40-core part. We store it per configuration precisely because publishing a single figure for a chip that ships in two meaningfully different forms would misrepresent half the machines sold under that name. Both numbers appear on every system page, in the leaderboard and in the comparison tool, and they are the two columns worth sorting before any other.
