Last updated: September 2026
Laptop listings now carry a third processor spec alongside the CPU and the GPU, usually with a number in trillions attached. Most buyers have no framework for judging it, and the marketing is not designed to supply one. This article explains what the chip is, what the number does and does not tell you, and how to work out whether any of it should affect what you buy.
For the hardware picks themselves, our AI hardware section covers the gear. This page is the explainer that should come first.
What an NPU Is
Microsoft's own developer documentation gives the clearest short definition, and it is worth quoting rather than paraphrasing:
NPUs are designed specifically to execute the deep learning math operations that make up AI models.
The same page describes the hardware class built around them:
Copilot+ PCs are a new class of Windows 11 hardware powered by a high-performance Neural Processing Unit (NPU) — a specialized computer chip for AI-intensive processes like real-time translations and image generation — that can perform more than 40 trillion operations per second (TOPS).
(Microsoft Learn, Windows AI documentation, checked 6 September 2026.)
The useful mental model is a division of labour between three processors. The CPU is a generalist: it does anything, one thing at a time, very fast. The GPU is a specialist in doing the same arithmetic to enormous numbers of values simultaneously, which is what both graphics and neural networks need. The NPU does a narrower version of that same parallel arithmetic, at much lower power, and it is designed to run continuously without draining the battery or spinning up a fan.
Which means the NPU's real advantage is rarely raw speed. A serious discrete GPU will usually beat an NPU on throughput. The NPU wins on performance per watt, and on being available for background work that would be absurd to run on a GPU all day.
What the TOPS Number Does and Does Not Tell You
TOPS is trillions of operations per second: a theoretical peak throughput figure. Two things about it are worth being precise about, because the marketing is not.
First, 40+ TOPS is a platform requirement, not a quality threshold. Microsoft's documentation states that "Many of the new Windows AI features require an NPU with the ability to run at 40+ TOPS". That is a floor for a feature set to be enabled, in the same way a game lists a minimum specification. It is not a statement that 45 is meaningfully better than 41, and it says nothing about anything outside that feature set.
Second, we could not find a published methodology for how vendors measure it. We looked, at source, and Microsoft's documentation states the requirement without stating how the figure is derived. That matters because peak-throughput numbers are normally quoted at a specific numerical precision, and a figure measured at lower precision is arithmetically larger for the same silicon. Absent a stated, common methodology, comparing one vendor's TOPS against another's is comparing two numbers that may not have been measured the same way. We are not going to tell you what the differences are, because we could not verify it — we are telling you the comparison is less solid than it looks.
Treat TOPS the way you would treat a car's peak horsepower: real, relevant, and a poor predictor on its own of whether the thing will do what you want.
The Constraint That Usually Bites First Is Memory
If your interest in an NPU is running AI models locally rather than sending data to a server, the specification that decides what is possible is generally not the NPU at all. It is memory — how much, and how fast.
A model has to fit before it can run. A machine with a fast NPU and modest RAM will simply refuse to load a large model, while a machine with a slower NPU and plenty of memory will run it, unhurriedly. Once it fits, the speed of text generation is often bounded by how fast the machine can move the model's weights through memory rather than by how fast it can multiply them. This is why memory bandwidth appears in serious discussions of local AI performance and almost never in laptop marketing.
The practical ordering for anyone buying a machine to run models on: memory capacity first, memory bandwidth second, the accelerator third. That is the reverse of how the specifications are advertised.
What Actually Uses the NPU Today
The honest answer is: a specific, growing, still fairly short list, and it does not automatically include the AI application you have in mind.
The workloads that suit an NPU are the ones that run continuously, in the background, and would otherwise cost battery: video call effects such as background blur and framing, live captioning and translation, noise suppression, on-device search indexing, and various operating-system features. These are exactly the tasks you want off the CPU and off the GPU.
The important caveat is that software has to be written for it. An application that has not been built to target the NPU will not use it, no matter how capable the chip is. Much of the current generation of local AI tooling targets the GPU, because that is where the mature software ecosystem is. This gap is closing, but "my laptop has an NPU" and "my AI application is faster" are two separate claims and the second does not follow from the first.
