whichaipc

Independent hardware bench · UK

Which machine will actually run your models?

Buying for local AI is a peculiar sort of shopping. The spec sheets are written for datacentre buyers and the benchmark videos rarely use the quantisation you'll run. So this is the plain version: what fits in the memory, how quickly it answers, and what it costs here in Britain.

Everything here runs past a dual RTX 4090 48GB bench before it gets written up, and where I've had to rely on someone else's numbers, the page says so.

On the bench

GPUs compared
15
Prebuilt machines
8
Largest VRAM covered
96 GB
Top GPU score
4/5
Top AI PC score
3.8/5

Scores are our own weighted rubric. VRAM and memory bandwidth carry the most weight, because between them they decide what loads and how fast it answers.

Route 01

Put a card in the PC you own

Nearly always the cheaper route, and the more upgradeable one. VRAM decides what you can load at all, bandwidth decides the pace, and a used 24GB card still does more for the money than anything newer.

Best GPUs £ per GB Multi-GPU Power caps
See the ranking
Route 02

Buy a machine built for it

Unified-memory boxes changed the maths. A mini machine with 128GB will hold models no consumer graphics card can touch, though it answers more slowly once it has them. That trade is the whole decision.

DGX Spark Strix Halo Mac Studio Prebuilt towers
Compare the machines

Instruments

Work it out before you spend

The question that decides everything is whether the model you want fits in the memory you're buying. You can answer that in about a minute.