Every AI laptop sold since 2024 has a number on it. 40 TOPS, 45, 48, 50, 60, 80, and on some marketing pages a triumphant 120. It is printed the way a clock speed used to be printed, as though bigger is simply better and the comparison is obvious.
It is not obvious, and the reason is not that anybody is lying. It is that TOPS is a family of numbers rather than one number, the vendor picks which member of the family to print, and the footnote that would make it comparable is usually somewhere else.
This guide covers what the unit actually measures, the three separate figures one chip legitimately has, what precision does to all of them, which one Microsoft's Copilot+ requirement is written against, and the five questions that turn a headline back into a specification.
What the unit measures
TOPS is trillions of operations per second. It is arrived at the way a factory quotes capacity: count the multiply-accumulate units, multiply by how many operations each does per clock, multiply by the clock. It is a ceiling, computed from the design rather than measured from a workload, and no real model will reach it.
That is fine as far as it goes. A CPU's clock speed is the same kind of number and nobody thinks a 5 GHz processor does five billion useful things a second. The trouble starts when the same word is used for several different ceilings on the same box.
Three numbers, one chip
Take Intel's Core Ultra 200V series, announced with figures Intel publishes itself: up to 120 total platform TOPS, made up of up to 67 GPU TOPS and up to 48 NPU TOPS.
None of those is wrong. They answer three different questions:
| Number | Answers | Useful for |
|---|---|---|
| Platform TOPS | What could the whole processor do if every engine ran flat out at once | Almost nothing you will experience. No single workload runs on all three engines at once. |
| GPU TOPS | What the integrated graphics can do on AI work | Real, and often the faster path for image generation and larger models. |
| NPU TOPS | What the dedicated neural engine can do | The one Windows feature gating is written against, and the one that runs at low power. |
So a laptop sticker reading "120 TOPS AI PC" describes a machine with a 48 TOPS NPU. Both statements are accurate. Only the second one tells you which features you get.
Precision doubles it, and doubles it again
Numeric precision is the second axis, and it moves the figure more than most people expect.
An accelerator that performs a fixed number of multiply-accumulate operations per clock at 8-bit integer precision generally performs about twice as many at 4-bit, because the same datapath carries two of the smaller numbers where it carried one. Some designs then quote a further doubling for structured sparsity, where half the weights are pruned to zero in a pattern the hardware can skip.
Same silicon, no exaggeration anywhere, and up to four legitimate figures with a fourfold spread between the smallest and the largest. Which one appears on the product page is an editorial decision.
This is not an abstract concern for anybody running local models, because Windows itself expects the small formats. Microsoft's own developer guidance for Copilot+ PCs puts it plainly: models are usually trained and distributed in larger formats such as FP32, "many NPU devices, however, only support integer math in lower bit format, such as INT8", and a model therefore has to be quantized before it can run on the NPU at all.
Which number the Copilot+ bar is written against
Microsoft's definition of a Copilot+ PC is an NPU that can perform more than 40 TOPS. It is written against the neural engine on its own, not the platform total, which is exactly why a machine can advertise triple digits and still sit outside the club, and why some machines with modest headline numbers are inside it.
The second half of the requirement gets less attention and rejects more machines: 16 GB of RAM. An 8 GB laptop with a qualifying NPU is not a Copilot+ PC, and there is no update that changes that, because the features are gated on the whole configuration rather than on the chip.
What that buys is a defined set of Windows features that run on the device: Windows Studio Effects, Live Captions with translation, Recall, Click to Do, Cocreator in Paint, super resolution in Photos. Whether your own applications use the NPU is a separate question, and the answer today is mostly no. Our guide on whether your NPU is doing anything covers how to check rather than assume.
What TOPS cannot tell you
Three things, and each of them decides more about your experience than the headline does.
How fast memory can feed it. Generating text from a language model reads the whole model out of memory for every token produced. That makes memory bandwidth, not arithmetic, the thing that sets the pace, and TOPS does not describe memory at all. A laptop NPU usually has no memory of its own and shares system RAM with everything else. The arithmetic in the memory guide is the other half of this picture.
What the toolchain supports. An NPU only runs models compiled for it, through that vendor's execution provider. Windows ML picks a provider at runtime and falls back to the GPU or CPU when the right one is missing, silently and correctly. A chip with more TOPS and thinner model support loses to a slower one with a mature stack on every model that only the slower one can run.
Sustained versus peak. TOPS is a peak. A thin laptop that reaches it for four seconds and then settles into a thermal limit has a different product than the number suggests. This is the same story as CPU boost clocks, and it is measured the same way: run the work for ten minutes and read the counters at the end, not at the start.
Five questions for a spec sheet
- Which engine? NPU, GPU, or a platform total that adds up things which never run together.
- At what precision? INT8 is the common reference point. INT4 roughly doubles the figure and does not double the capability.
- With sparsity or without? If a footnote says "with sparsity", the dense number is about half of what is printed.
- How much memory, and how fast? Capacity decides which models fit; bandwidth decides how quickly they answer.
- Which runtime supports it? A chip is only as useful as the models somebody has shipped for it.
Answer those five and you have a specification. Answer none of them and you have a sticker.
Where PC Workman stands on this
PC Workman does not read NPU counters and does not report a TOPS figure. It reads what can be measured on a running machine: processor and memory load, temperatures, power, storage and process behaviour. The reason it does not print a TOPS number is the reason this guide exists. That figure comes off a specification sheet rather than off the hardware, and a monitor that repeats a marketing number back to you has not measured anything.
What it can tell you is the part that is real on your own machine: whether the work you expected to be offloaded is actually being carried by your CPU, what that costs in heat, and whether the chip is holding its clocks or sitting in a thermal limit.
Next question: how much memory does a local model actually need? Browse the rest of the guides →
Sources: Intel, Core Ultra 200V series announcement (up to 120 platform TOPS, up to 67 GPU TOPS, up to 48 NPU TOPS); Microsoft, Copilot+ PCs developer guide (the 40+ TOPS NPU definition, and quantization to integer formats such as INT8); Windows ML execution providers (runtime provider selection and fallback).