VRAM, RAM or unified memory: which one does an LLM use?
Where the model has to sit on a PC with a graphics card, a PC without one and a Mac.
Short answer
The memory of whatever runs it: VRAM on a graphics card, unified memory on a Mac, system RAM on a PC without a card.
- Graphics card: Only its own VRAM. An RTX 4090 offers 24 GB, whatever the PC's RAM.
- Mac: Unified memory, but macOS lets the GPU use only part: about 104 GB on a 128 GB Mac.
- PC without a card: System RAM. It works, but more slowly.
Computed with the site's fit rule and Apple's Metal working-set limit. Methodology
Three kinds of memory
A model runs where its weights are loaded. What counts is the memory the processor doing the work can reach quickly, not the total printed on the box.
- VRAM on a graphics card
- Memory soldered on the graphics cardThe PC's RAM does not add to it. The site reserves 2 GB per card, because the system runs from its own RAM.
- System RAM
- The computer's main memoryUsed when there is no graphics card, or by layers that do not fit in VRAM. The site reserves 25% of it for the system and apps, at least 2 GB and at most 8 GB.
- Unified memory
- One pool shared by the processor and the GPUMacs with Apple silicon. Large, but macOS caps the GPU's share. The site also reserves 25% of it (2 GB to 8 GB).
Speed follows the same split: every word the model writes reads its active weights from memory once, so memory bandwidth (GB/s) is a large part of the speed. Hardware pages list each machine's bandwidth; the site shows tokens per second only when someone measured that exact machine (fitting is not speed).
The macOS limit
What each Mac offers to a model under the site's rule.
Metal, the macOS graphics layer, reports a recommended working-set size: how much memory the GPU should use without hurting the system (Apple · recommendedMaxWorkingSetSize, read Oct 17, 2026). Apple does not publish a fixed share; the site plans with 75% of memory, a value commonly reported on recent Macs that can be lower on some macOS versions and small machines, plus its system reserve, and never with more than the Mac has:
- Mac mini M4 · 16 GB unified
- Offers 16 GB to a modelof 16 GB installed: the system reserve is the larger cut
- Mac mini M4 · 24 GB unified
- Offers 24 GB to a modelof 24 GB installed: the system reserve is the larger cut
- Mac mini M4 Pro · 48 GB unified
- Offers 44 GB to a modelof 48 GB installed: the GPU share is the limit
- Mac Studio M5 Max · 128 GB unified
- Offers 104 GB to a modelof 128 GB installed: the GPU share is the limit
Advanced users can raise the GPU limit in Terminal until the next restart; macOS still needs room for itself (how, in Local AI basics).
One model, four machines
A 27B model in Q4, short conversations. Planning estimates in whole GB.
| Machine | Memory kind | 27B Q4 needs (GB) | Offers to a model (GB) |
|---|---|---|---|
| RTX 4090 · 24 GB VRAM | VRAM on a graphics card | 23 | 24 |
| Dell XPS 14 (2026, DA14260) · Core Ultra 7 358H · Intel Arc Graphics · 32 GB | System RAM | 29 | 32 |
| Mac mini M4 Pro · 48 GB unified | Unified memory | 29 | 44 |
| Mac Studio M5 Max · 128 GB unified | Unified memory | 29 | 104 |
When it does not fit
llama.cpp can split a model between the graphics card and the processor when it is larger than the VRAM (llama.cpp README, read Oct 17, 2026). Ollama and LM Studio are built on llama.cpp (Hugging Face · Ollama, read Oct 17, 2026) (Hugging Face · LM Studio, read Oct 17, 2026). The model runs, but much more slowly, because the layers left in system RAM are read from slower memory on every word. Hardware pages list what fits in VRAM, the fast case.
On a Mac there is no second pool to spill into: the GPU and the processor share the same memory, so the model has to fit under the macOS limit, or the limit has to be raised.