What Is Unified Memory? How Macs Use It as VRAM for Large AI Models

Run a large language model on a typical gaming PC and you hit a wall fast: the graphics card has maybe 12GB or 16GB of dedicated VRAM, and once the model exceeds it, performance falls off a cliff. Unified memory sidesteps that wall entirely. On Apple Silicon Macs, the CPU and GPU draw from one shared pool, so the graphics side can reach far more memory than any consumer GPU carries on its own.

Quick Answer

Unified memory is a single pool of RAM that the CPU and GPU access together, with no separate VRAM and no copying between them. On Apple Silicon, that means the GPU can use almost the whole memory bank, so a Mac with 128GB unified memory can hold a 70B-parameter model (quantised to Q4_K_M) entirely in memory at once.

What Unified Memory Replaces

A normal PC keeps two separate memory stores: system RAM for the CPU and dedicated VRAM on the graphics card. When the GPU needs data the CPU is holding, the system copies it across the PCIe bus first. That copy costs time and caps how big a model you can load by the GPU's VRAM alone.

Apple Silicon puts a single memory bank on the same package as the chip. Both the CPU and the GPU address it directly. Nothing gets duplicated, and the GPU is no longer boxed in by a small dedicated allocation. If you want the same trick on the Windows side, the new wave of AI-focused desktops uses comparable shared-memory architectures, and the dedicated AI PC lineup at Evetech is the place to see which models lean into large local memory pools.

Why This Matters for Local AI Models

Model size is measured in parameters, and bigger models need more memory to load. A 70B model quantised to a 4-bit format needs roughly 40GB or more just to sit in memory before it generates a single word.

On a conventional GPU with 16GB of VRAM, that model simply will not fit, so part of it spills to slower system RAM and inference crawls. A Mac configured with 128GB of unified memory keeps the entire model resident in fast memory and lets the GPU work on it directly. That is the difference between a model that runs and one that stalls.

The trade-off is bandwidth. Unified memory is shared, so heavy CPU and GPU workloads compete for the same pool, and a top-tier discrete GPU still pushes more raw memory bandwidth for the data it does hold. For loading and running large models locally, though, having the capacity at all usually beats having faster but far smaller VRAM.

Who Should Care

If you are experimenting with running models on your own machine rather than calling a cloud API, memory capacity is the spec that decides what you can do. Choosing an Apple Silicon Mac (or a Windows AI PC built around a large shared pool) with as much unified memory as your budget allows is the single most useful decision for local inference. To gauge what local buyers are pairing with this kind of work, the current best-selling desktops are a quick reality check on real-world configurations.

Frequently Asked Questions

Does unified memory count as VRAM?

Functionally, yes. On Apple Silicon the GPU draws from the shared pool, so a large unified-memory configuration gives the GPU access to far more memory than a typical discrete card carries as dedicated VRAM.

Can I run a 70B model on a Mac?

With enough unified memory you can. A 70B model at a 4-bit quantisation needs roughly 40GB or more, so a 128GB Mac holds it comfortably while leaving headroom for the operating system and other apps.

Is unified memory faster than a dedicated GPU?

Not in raw bandwidth. A high-end discrete GPU moves data faster within its own VRAM. Unified memory wins on capacity and on skipping the copy step, which is what large local models actually need most.

How much unified memory do I need for AI work?

For smaller models 16GB to 32GB is workable. Serious local inference on large models wants 64GB or more, and 128GB opens up the biggest models you can practically run at home.