Unified memory: Apple and Strix Halo

Most PCs have a wall between the CPU's RAM and the GPU's VRAM. The GPU can only use its own VRAM, so the model has to fit in the card, or you offload and take a speed hit. Unified memory removes that wall: one pool that CPU and GPU both use.

Why it changes everything

With unified memory, the GPU can address the whole system RAM, not just a few gigabytes of dedicated VRAM. A Mac or a Strix Halo laptop with 64 GB or 128 GB can run models that would need a stack of discrete cards, because the model and the KV cache all live in that one pool.

That is why "128 GB" on an Apple machine is not comparable to "24 GB VRAM" on a PC. The Apple machine runs a 70B comfortably; the 24 GB card does not.

The two catches

1. Bandwidth, not capacity, is the limit. Unified memory is usually slower than the GDDR memory on a discrete card, so tokens-per-second is lower than the "it fits" number suggests. A big model on a Mac fits, but runs at a laptop's pace.
2. You are sharing. The OS, your apps and the model all draw from the same pool. A 64 GB machine does not have 64 GB free for the model; it has whatever is left after everything else.

When unified memory wins

• You want to run models bigger than any card you can afford.
• You want long context without watching a VRAM budget.
• You value quiet, low power and simplicity over raw throughput.

When a discrete card wins

• You want maximum tokens-per-second on a model that already fits in VRAM.
• You need CUDA for a specific tool or for fine-tuning.
• You run a server with concurrent users, where a real GPU's bandwidth and batching matter.

The decision

Unified memory is about capacity and convenience; a discrete GPU is about speed and the NVIDIA ecosystem. Know which one you are buying before you compare their numbers, because they are not the same race.

The catalog has both classes clearly separated. Filter by the machine type and you will see the two worlds side by side, with their real reported speeds.