What Is an NPU, and Do You Need One?
Updated
The NPU is the most marketed and least understood component in the current generation of laptops, and the gap between what it does and what buyers assume it does is wide.
What is an NPU?
A neural processing unit is a small, extremely power-efficient accelerator designed for sustained, low-intensity inference. It runs small quantised models continuously without draining the battery, which is exactly what live captions, background blur, translation and local search need.
Does it make local models faster?
Essentially no. When you run a language model through Ollama or LM Studio, the work goes to the graphics processor or the processor, not the NPU. A 50 TOPS neural processor and a 10 TOPS one will run the same 32B model at effectively the same speed, because neither is doing the work.
So why does every laptop advertise one?
Because Copilot+ certification requires at least 40 TOPS, and that badge sells laptops. The certification is a real thing and the features are genuinely useful. It simply measures something different from the capability most people buying an AI PC think they are buying.
When should the NPU affect your purchase?
If you want the Windows AI features to run well on battery, buy for the NPU and ignore everything else on this site. If you want to run models yourself, ignore the TOPS figure entirely and buy memory and bandwidth. Those are different purchases and conflating them is the most expensive mistake in this category.
What does the NPU actually run?
Small, heavily quantised models that need to run continuously without draining the battery. Live captions transcribing audio as it arrives, background segmentation in a video call, on-device translation, local semantic search over your files, and image cleanup. These are real features that people use daily, and the NPU is the right hardware for them.
The common thread is that they are small, constant and latency-sensitive. That is a different problem from generating text with a 32B model, which is large, bursty and throughput-sensitive.
Why can't it run a large model?
Two reasons. Memory, first: the NPU draws from the same shared system memory as the integrated graphics, so it faces the same ceiling, typically about half of system RAM. Second, support: the runtimes people actually use to run local models target the graphics processor or the processor, and NPU paths are limited, vendor-specific and immature.
This may change. Vendors are working on it and the tooling is improving. It is not a reason to pay more for TOPS today.
Will this change?
Probably, gradually. If NPU inference paths mature and memory allocation improves, these chips could handle small models efficiently and extend battery life meaningfully. That would be a genuine benefit, and it would still not turn a shared-memory laptop into a machine that runs a 70B model, because the memory ceiling is a separate constraint from the accelerator.
What should you actually buy?
If the Windows AI features matter and you work on battery, buy a Copilot-class machine and choose on screen, keyboard and weight. If you intend to run models yourself, treat the TOPS figure as noise and spend on memory and bandwidth. The table above lists every machine we track with an NPU, with its usable AI memory alongside, so you can see directly how little the two numbers relate.
How the NPU affects battery life
Meaningfully, and this is its genuine advantage. Running background transcription or video effects on the graphics processor is a measurable drain; running them on a neural processor designed for the job costs a fraction of the power. On a thin laptop that can be the difference between a full working day and an afternoon.
If you take video calls all day on battery, that benefit is real and worth paying for. It is simply a different benefit from the one the term AI PC implies to most buyers.
What memory do these machines actually give a model?
About half of system memory, because integrated graphics share the pool. A 16GB machine gives a model roughly 8GB, a 32GB machine roughly 16GB. The first runs a small model, the second a mid-sized one at a short context.
On most Copilot-class laptops the memory is in the processor package and cannot be upgraded, so this ceiling is permanent for the life of the machine. That makes the configuration you choose at purchase unusually important.
The one-sentence version
The NPU makes your laptop better at the AI features Microsoft ships, and makes essentially no difference to running models yourself. Buy for it if you want the former and ignore it entirely if you want the latter.
How we calculate the figures on this page
The distinction this page draws is visible in the data rather than being an opinion about marketing. We record two separate figures for every machine: NPU throughput in TOPS, and usable AI memory in gigabytes. On a Copilot-class laptop the first is typically 45 or 48, which clears Microsoft's certification bar comfortably, while the second is about half of system memory, so roughly 16GB on a 32GB machine. A 32B model at Q4 needs about 24GB. The machine therefore passes the AI PC certification and cannot run the model, and both of those statements come from its own recorded specification rather than from a judgement about the category. Our AIPC Score weights NPU capability at 5 percent of the overall ranking and AI memory at 20 percent, with model compatibility at a further 15, which is why these machines rank where they do. That weighting is published and the component breakdown prints on every system page, so if you disagree with our emphasis you can see exactly what we weighted and judge the machine on the components that matter to you instead. The table above lists every machine we track that carries an NPU, with its TOPS figure and its usable AI memory side by side.
