Unified memory is one pool of RAM that the CPU and GPU share, instead of the GPU having its own separate video memory. It is why a thin laptop with 128 GB of unified memory can load an AI model that would not fit on a $2,000 graphics card with 32 GB. Apple has used it in every Apple silicon Mac, and NVIDIA's RTX Spark laptops brought it to Windows on October 16, 2026. Here is how it works, what it is good and bad at, and how much you need.
How unified memory works
In a typical desktop PC, the processor and the graphics card each have their own memory:
- System RAM sits next to the CPU, often 16 to 64 GB.
- VRAM sits on the graphics card, usually 8 to 32 GB, and is very fast.
- Data moves between them over the PCIe slot, which is much slower than either memory.
With unified memory, the CPU and GPU are in the same chip package and read the same RAM. NVIDIA's RTX Spark porting guide describes up to 128 GB shared by the CPU and GPU over a 256-bit interface at 300 GB/s, and calls the memory "fully coherent": both processors see the same data without copying it back and forth. Apple describes each M-series chip as one system on a chip with a unified memory architecture.
Integrated graphics in ordinary laptops also share system RAM, so the idea is not new. What is new is pairing shared memory with a GPU powerful enough, and memory fast and large enough, to compete with dedicated graphics cards for some jobs.
Unified memory vs VRAM
| Unified memory | Dedicated VRAM | |
|---|---|---|
| Who uses it | CPU and GPU share one pool | GPU only |
| Typical size | 16 to 128 GB in laptops (512 GB in Apple's Mac Studio with M5 Ultra) | 8 to 32 GB on consumer cards |
| Bandwidth | 300 GB/s (RTX Spark) to 614 GB/s (M5 Max) | Often higher on desktop graphics cards |
| Copying between CPU and GPU | Not needed | Over PCIe |
| Upgradable | No, soldered to the chip | Only by replacing the card |
The sizes for Apple and NVIDIA come from Apple's M5 Max announcement, its Mac Studio announcement, and NVIDIA's RTX Spark specifications.
The trade-off in one sentence: unified memory wins on size, dedicated VRAM usually wins on speed. A graphics card's memory is built only for the GPU and is generally faster per gigabyte, but there is less of it, and the card cannot borrow more without a big slowdown.
Why unified memory matters for local AI
Running a large language model on your own computer needs two things:
- Enough memory to hold the model. The weights take roughly parameters × bits per weight ÷ 8. A 70-billion-parameter model at 4.8 bits per weight (the common Q4_K_M format) needs about 42 GB, plus space for the context. No consumer graphics card holds that on its own, but a 64 GB or 128 GB unified memory machine does.
- Enough bandwidth to read it quickly. For each word-piece (token) it generates, a dense model reads all of its weights once. So the theoretical top speed is the memory bandwidth divided by the model's size in memory.
Working through the second point with the official bandwidth figures:
| Machine | Bandwidth | 70B dense at Q4_K_M (42 GB) | 120B mixture of experts, 5B active, 4.5-bit (about 2.8 GB read per token) |
|---|---|---|---|
| RTX Spark laptop | 300 GB/s | about 7 tokens/s ceiling | about 100 tokens/s ceiling |
| M5 Max MacBook Pro | 614 GB/s | about 15 tokens/s ceiling | about 220 tokens/s ceiling |
These are upper bounds; real speeds are lower. The table shows why mixture-of-experts models suit unified memory machines: they need a lot of memory to hold every expert, but read only a small part of it per token. Our mixture of experts guide explains why. Plug your own model and machine into the LLM VRAM Calculator: enter the unified memory as "GPU VRAM" and the bandwidth to see both the fit and the speed ceiling.
How much unified memory do you need?
Because unified memory cannot be upgraded later, choose for the next few years of use:
- 16 GB: browsing, office work, photo editing and small AI models up to about 8B parameters.
- 32 GB: video editing, software development with containers, and models up to about 14B to 32B parameters at 4-bit.
- 64 GB: 70B-class models at 4-bit with moderate context, or several large creative apps at once.
- 128 GB: 100B-class mixture-of-experts models, long contexts, or running a model while doing other heavy work.
Leave headroom. The operating system and your open apps use the same pool, so a 128 GB machine cannot give all 128 GB to the GPU. If a model barely fits, a shorter context or an 8-bit KV cache frees memory before you drop to a lower-quality quantization.
To see how much memory a file will take, the data storage converter turns GB into GiB, which is how some tools report sizes.
Does unified memory help games?
Not much. Games are written for graphics cards with 8 to 16 GB of VRAM and rarely need more, while they do benefit from the very high bandwidth of dedicated cards. A laptop with unified memory plays games well when its GPU is fast enough, as NVIDIA claims for RTX Spark, but the large memory pool itself gives little advantage. It matters most for AI models, 3D scenes with huge textures, and scientific data that would not fit in VRAM.
How to check what your computer has
- Windows: open Task Manager (Ctrl+Shift+Esc), choose Performance, then GPU. "Dedicated GPU memory" is VRAM; "Shared GPU memory" is system RAM the GPU can borrow, which is slower on PCs with a separate graphics card.
- Mac: Apple menu → About This Mac shows the chip and the memory, all of it unified.
- Any browser: What GPU Do I Have? shows which graphics chip your browser uses and whether it is integrated or dedicated.
Common questions
Is unified memory the same as RAM? Yes, it is ordinary system RAM. The difference is that the CPU and GPU share all of it directly instead of the GPU having separate video memory.
Is unified memory better than VRAM? It is larger and avoids copying data, which helps big AI models. Dedicated VRAM is usually faster, which helps games and smaller models. Which is better depends on the job.
Can I upgrade unified memory later? No. It is part of the chip package on Apple silicon Macs and RTX Spark laptops, so buy the amount you will need.
Is 128 GB of unified memory enough to run a 120B model? A 120B mixture-of-experts model at about 4.5 bits per weight needs around 68 GB for its weights, so it fits with room for context. NVIDIA says RTX Spark runs models up to 120B parameters.
Does a Mac have more memory bandwidth than RTX Spark? Apple lists up to 614 GB/s for the M5 Max, against 300 GB/s in NVIDIA's RTX Spark documentation. Apple has not published the details behind that figure, so compare real benchmarks once they appear.