2026 · 10 · 11 07:53

Mac Unified Memory vs RTX 5090 VRAM for Local LLMs: What to Buy

Is Mac unified memory really VRAM? RTX 5090 has 32 GB at 1,792 GB/s; Mac Studio M5 Ultra goes to 512 GB at 1.2 TB/s. A one-minute formula to see which machine runs your model, plus local hardware vs API cost.

Open on YouTube
Watch on
The original Chinese episode Mac 统一内存是内存还是显存？5090、DGX 怎么选

What this video covers

Chapters

Transcript

$2,000 card, 32GB. Mac Studio, 512GB?

A $2,000 graphics card ships with 32 GB of memory. A Mac Studio can be configured with 512. Apple calls it unified memory: the CPU and the GPU share 1 pool of memory. Is that memory actually VRAM? And should you buy a Mac, a card, or that weird little NVIDIA box? By the end, you’ll be able to do the math yourself in under 1 minute.

RAM vs VRAM: two pools vs one

First, how a normal PC works. Your RAM is 1 pool. Your graphics card carries its own separate memory, called VRAM. It sits on the card, right next to the chip that does the work. And that chip works fast only on data inside its VRAM. So the model has to be copied over first, through a connection called PCI Express. Think of it as a shipping lane between 2 warehouses. Apple took a different approach. On a Mac, the CPU and the GPU sit on 1 chip and reach into the same pool directly. No copying, no shipping lane. Everything sits on the same shelf. So unified memory is both RAM and VRAM: 1 pool of memory doing both jobs. That’s a real advantage. A Mac can feed the GPU a model that would never fit on a card.

The macOS GPU limit

But hold on. macOS doesn’t let the GPU use all of it. By default, the GPU’s share is about 2/3 to 3/4 of your total memory. The rest stays reserved for the system. On a 24 GB Mac, a model can really use about 16 to 18 GB, not 24. That’s the most common mistake people make when shopping.

Capacity: will the model fit?

From here on, you only need 2 numbers. Capacity: how much it can hold. And bandwidth: how fast it can move. Capacity decides whether the model fits at all. Quick rule of thumb. The B in a model’s name means billions of parameters. A common 4-bit version, called quantized, takes roughly that number × 0.6 GB. Mistral Small, 24B, needs about 15 GB. Qwen3 at 32B, about 20. gpt-oss-120b, around 65. And the full DeepSeek, 671 billion parameters, wants about 404 GB. Big numbers. Now let’s use them.

Bandwidth: how fast it talks

Bandwidth decides how fast it talks. A language model writes 1 word at a time, and each new word means reading through the model’s weights again. Wider road, faster words. Here are the numbers. MacBook Air: 153 GB per second. The DGX Spark: 273. The RTX 5090: 1,792. And the Mac Studio with M5 Ultra: about 1.2 TB per second. That’s roughly a third below the 5090.

The RTX 5090’s 32GB wall

So the 5090 wins on speed, but it has 1 hard limit: 32 GB, full stop. If your model fits, it flies. The moment it doesn’t, part of the model has to live in system memory and commute over that shipping lane for every single word. Speed doesn’t dip. It falls off a cliff, several times slower, sometimes over 10×. Same idea as rush hour. 1 more thing to be honest about. If you want to feed it long documents, or generate images and video, now you’re fighting over raw compute and the CUDA ecosystem. CUDA is NVIDIA’s software platform. Most image and video tools are only mature on CUDA. That round isn’t close.

Which machine is yours?

So which machine is yours? 4 kinds of buyers. Find yourself. Type 1: your computer is for everyday work, and AI is for chatting, writing, and code help. My advice: buy memory for your normal needs and don’t pay extra for models. My own laptop is a 16 GB MacBook Air, and it’s getting tight. But since I don’t run models locally, I’d upgrade to 24 at most. 24 gigs can technically fit a Mistral Small, but it’s bare-minimum tight. And small models are a real step down anyway. Type 2: you want to play with models at home, in the teens to 32B range. Then 32 GB is the floor. Don’t haggle with yourself. A Mac mini with M6 tops out at 32. The M5 Pro version goes to 64. Tiny box on your desk, done. Type 3: you’re making images and video, or you’re fine-tuning. That means teaching a model with your own data. Or you’re writing CUDA code. Stop deliberating: desktop PC with a 5090. 32 gigs of VRAM and a 575-watt card. Budget the power supply and the electric bill. Type 4: you’re set on running 100-billion-parameter models at home. 2 options. The DGX Spark: 128 gigs of unified memory, 273 GB per second. NVIDIA says it handles roughly 200B models, priced just under $5,000. Or the Mac Studio with M5 Ultra: up to 512 gigs, 1.2 TB per second. But the Ultra starts at $5,500, and the 512 upgrade costs a fortune on top. And even then, full DeepSeek won’t fit by default: 404 gigs is more than the 3/4 macOS hands to the GPU.

Do the math yourself

Last thing: learn to do the math yourself. Open “About This Mac” and multiply your memory by 2/3. That’s a conservative estimate of what a model can use. Then take the model’s parameter number and multiply by 0.6. Compare the two. You’ve just figured out in 1 minute what most people learn after buying the wrong machine. One optional power-user tip. There’s an undocumented terminal command, it’s on screen now, that raises the GPU’s memory ceiling. It resets on every reboot, and Apple doesn’t document it. Starve the system and the machine can crash. Leave real headroom, and treat it as a tweak, not free memory.

Local hardware vs API

Now 1 last bill. DeepSeek’s official API, meaning you pay per use, charges its Flash model 60 cents per million output tokens at off-peak hours. A token is roughly a chunk of a word; it’s the unit AI bills by. Take the price of a 5090 and convert it. At the off-peak rate that’s about 3 billion output tokens. Even at the peak rate it’s more than 1.5 billion, and far more if most of your traffic is input. If your usage is light, an API plus a good subscription beats new hardware by a mile. Here’s your next step. Before you buy anything, run the 2-line calculation: your memory × 2/3, the model’s size × 0.6. If they don’t fit, don’t force it. That’s what the API is for. I keep doing this math on AI gear and subscriptions so you don’t have to. Subscribe, and the next bill lands in your feed. And now you tell me. How much memory does your machine have: 16, 32, or more? Have you tried running a local model? Drop your number in the comments.

FAQ

Is Mac unified memory the same as VRAM?

Yes, in practice it is both. On Apple Silicon the CPU and GPU sit on one chip and share one pool of memory, so the GPU reads the model directly with no copy. The catch: by default macOS lets the GPU use only about two-thirds to three-quarters of it.

How much memory do I need to run a local LLM?

A common 4-bit model needs roughly its parameter count in billions × 0.6 GB: Mistral Small 24B about 15 GB, Qwen3 32B about 20 GB, gpt-oss 120B about 65 GB. On a Mac, multiply your memory by two-thirds to see what a model can use, then compare.

Mac Studio or RTX 5090 for local AI?

If your model fits in 32 GB, or you do image, video or fine-tuning work, the RTX 5090 is faster (1,792 GB/s) and CUDA tools are more mature. For large models that need more memory, a Mac Studio with M5 Ultra goes to 512 GB at 1.2 TB/s; even then full DeepSeek (404 GB) does not fit in the GPU’s default share.

Sources

Mac unified memoryRTX 5090VRAMlocal LLMMac StudioDGX Sparkmemory bandwidth