You bought a Copilot+ machine, or a laptop whose box says AI PC, and somewhere in it is a neural processing unit rated at 40, 50, 60 or 80 TOPS. You open Task Manager to watch it work. It sits at 0%. You open it again during a video call, while a chatbot answers, while an image generates. Still 0%.
The obvious conclusion is that something is broken, and it is almost always wrong. An idle NPU is the normal state of an NPU. It is not a general accelerator that Windows spreads work across the way it balances CPU cores. It runs models that were converted and compiled specifically for it, which means a program uses your NPU only if somebody shipped a model built for that exact hardware. Most did not.
This guide covers where the activity shows up, which parts of Windows genuinely use it today, the three reasons the number stays at zero, and how to record a trace that answers the question instead of guessing at a graph that updates once a second.
Where NPU activity actually shows up
There are three places, and two of them are switched off until you switch them on.
The graph. Task Manager, Performance tab. On a machine with an NPU there is an entry under CPU, Memory, Disk, Wi-Fi and GPU, and clicking it gives utilisation, driver version and hardware location, the same shape of panel the GPU gets.
The per-process columns. This is the part most people never find. In the Processes, Users and Details tabs there are NPU and NPU Engine columns, and they are hidden by default. Right-click any column header, choose Select columns, and tick them. The NPU column gives the share of activity per process, and NPU Engine names which engine inside the NPU is running, which is how you tell two workloads apart on the same device.
The memory columns. The Details tab also offers NPU Dedicated Memory and NPU Shared Memory. On most consumer laptops an NPU has no memory of its own and borrows system RAM, so shared memory is where the number appears, and that fact matters later when somebody tells you a 40 TOPS chip will run a large language model.
What uses the NPU today
Microsoft gates a specific set of Windows features on an NPU capable of more than 40 TOPS, which is the definition of a Copilot+ PC rather than a marketing line. The features that reach the NPU on those machines include Windows Studio Effects, Live Captions with translation, Recall, Click to Do, Cocreator in Paint and super resolution in Photos.
Outside Windows itself the list is short and worth knowing, because it tells you when to look:
- Video calls. Background blur, framing, eye contact and noise suppression through Windows Studio Effects are the single most reliable way to make the number move. Start a call, turn the effects on, and watch the graph.
- Browsers. Microsoft Edge uses the NPU for some on-device work such as text recognition in PDFs.
- Apps that shipped their own model. A handful of creative and conferencing tools have NPU builds. Almost everything else runs its AI on the CPU, the GPU, or somebody's server.
Notice what is not on that list. A chatbot in a browser tab is a network request. A cloud image generator is a network request. Neither will ever touch your NPU, no matter how many TOPS it has.
Three reasons it reads zero
When you expected work and got a flat line, it is one of these three, in this order of likelihood.
1. Nothing asked for it. The default case, and not a fault. Windows does not migrate arbitrary work onto an NPU; the application chooses.
2. The model was in the wrong format and something else took the job. Models are trained and distributed in floating point formats such as FP32, and many NPUs execute integer math such as INT8, so a model has to be quantized and compiled for that hardware before it can run there. When the right execution provider is not present, Windows ML falls back to the GPU or CPU and the work completes normally. Nothing errors. Nothing tells you. The output is identical, the power draw is not.
3. The work was real but too short to see. Task Manager samples about once a second. A single image classification or a wake-word check can be over in a few milliseconds. Twenty of those in a minute is genuine NPU use that a one-second sampler will draw as zero for most of that minute.
Proving it with a trace
Windows ships an event tracing profile for exactly this. Windows Performance Recorder gained a Neural Processing profile that records the driver model's interactions with the NPU, so you get every submission with the process behind it rather than a sampled percentage.
From an administrator command prompt:
wpr -start NeuralProcessing -start CPU
:: now reproduce it: join a call with Studio Effects on, run your model, use the feature
wpr -stop npu.etl -compress
Open npu.etl in Windows Performance Analyzer, which is a free download from the Microsoft
Store, and look under Neural Processing for NPU Utilization. You get which processes submitted
work, when, and the callstacks that submitted it. If the trace is empty, the answer is settled: nothing
asked the NPU to do anything during that window.
Both tools come from the Windows Performance Toolkit, part of the Windows ADK. This is the same instrument Microsoft points its own developers at, which is worth knowing because it means the answer you get is the same one the people who wrote the driver would get.
What this means if you are buying
Three things follow from all of the above, and they are worth more than any TOPS comparison.
The NPU is a battery and thermals feature first. Where it wins is running a small model continuously at low power while the CPU and GPU stay free and quiet: background blur for the length of a meeting, captions for the length of a lecture. That is a real benefit and it is not the benefit the marketing implies.
It is not a substitute for GPU memory. An NPU on a laptop usually has no memory of its own and borrows system RAM. TOPS describes arithmetic rate, and the thing that limits a local language model is how fast weights can be read, so a large model does not become practical because the chip on the box got a bigger number. Our guide on how much VRAM you actually need covers the memory side.
The gated features need more than the chip. The Copilot+ bar is an NPU above 40 TOPS and 16 GB of RAM. A machine with a capable NPU and 8 GB does not qualify, and no update changes that.
Where PC Workman stands on this
Plainly: PC Workman does not read NPU counters today. It reads CPU, memory, storage, temperatures, power and process behaviour, and its assistant runs locally without an NPU and without a cloud key. NPU telemetry is a candidate for the roadmap rather than a feature I can claim, and a monitoring tool that invented a number for hardware it does not read would be worth less than no number at all.
What it does help with is the part of this that is measurable today: whether the work you think is on the NPU is actually being paid for by your CPU, and what that costs you in heat and battery. If the fans spin up during a video call on a machine with a 40 TOPS NPU, something is running in the wrong place, and that is visible in the counters PC Workman already reads.
Next question: what does a local model actually need from your machine? Browse the rest of the guides →
Sources: Microsoft, Copilot+ PCs developer guide (40+ TOPS requirement, Task Manager NPU view, quantization to integer formats, WPR Neural Processing profile); Windows ML execution providers (provider selection and fallback); Windows Performance Toolkit v11 (NPU analysis in WPA).