How we rank AI PCs
Every ranking on this site comes from the AIPC Score. It is not an editorial opinion and it is not a star rating. It is ten measured components, each normalised to a 0 to 100 scale with a published rule, combined using weights that depend on what you want the machine to do. The component breakdown appears on every system page and in the leaderboard row expander, so the number is always explainable.
The ten components
- AI memory. Usable memory for a model, on a logarithmic scale where 8GB scores 0 and 512GB scores 100.
- AI compute. FP16 throughput against a 250 TFLOPS reference.
- Memory bandwidth. Measured against a 1,800 GB/s reference, which is where current top-end discrete memory sits.
- Model compatibility. The share of a fixed reference set of models that fit at Q4 without compromise.
- GPU capability. Compute and bandwidth together for discrete cards, compute alone for unified-memory systems.
- NPU capability. Published TOPS, capped at 100.
- Software support. How well the local AI stack runs: CUDA scores highest, then Apple MLX and Metal, then ROCm and Vulkan, then shared integrated graphics.
- Power efficiency. Usable AI memory per watt.
- Expandability. Upgradeable memory, PCIe slots, spare storage bays and external GPU connectivity.
- Price and value. Usable AI memory per dollar, using the lowest verified offer. A system with no verified price scores zero here and is labelled as such.
Weights change with the workload
A machine that is excellent for image generation can be mediocre for a 70B language model, so one ranking cannot serve both. The Overall ranking weights AI memory and compute at 20 percent each, model compatibility at 15, and bandwidth and GPU at 10 each. The Local LLM ranking moves most of that weight onto memory, bandwidth and model compatibility, because those are what decide whether a model loads and how fast it generates. The Image AI ranking does the opposite and leans on raw compute and GPU capability.
What we will not do
We do not rank by commission. We do not publish a tokens per second figure unless it comes from a recorded benchmark with a source. We do not invent a discount percentage, and we do not publish review scores or star ratings, because we have not run a controlled review programme and pretending otherwise would be dishonest. Where a machine has not been benchmarked, the page says so.
When rankings change
Scores are recomputed whenever a specification is corrected or a price moves, because price is a scored component. A machine can therefore climb or fall without any hardware change. The leaderboard carries the timestamp of its last recomputation for exactly that reason.