Does an NPU actually save battery?

Search for whether an NPU saves power and you will find the same shape of claim everywhere: a few watts against a hundred and fifty, or ten times the efficiency, or a fifth of the consumption. The numbers vary and the structure does not, because most of them compare a neural engine in a laptop with a graphics card in a desktop.

Those are not two ways of doing the same job. They are two different machines with two different power budgets doing two different workloads. The comparison that decides anything for you is narrower and much easier: your machine, your task, with the NPU path available and then not.

This guide is a method rather than a verdict. Windows already ships the tools, the procedure takes about twenty minutes, and at the end you have a number about your own laptop instead of somebody else's slide.

Why the saving is real, and where it is not

The mechanism is not mysterious. A neural engine is fixed-function hardware built for one shape of arithmetic, running a model quantized to small integers. A general processor doing the same multiplication burns energy on everything that makes it general: fetching and decoding instructions, speculating, reordering, moving data through caches designed for other work. Specialised silicon doing one job does less of all that, and the difference shows up as watts.

So the saving exists. What the marketing tends to leave out is its shape:

• It applies to sustained small work. Background blur for an hour of meetings, captions for a lecture, a wake word listening all day. Long duration, modest model, no deadline. This is the case the hardware was designed for and where it wins clearly.
• It shrinks as the model grows. Once the job is limited by reading weights out of memory rather than by arithmetic, the engine doing the arithmetic matters less, because everything is waiting on the same memory. The memory arithmetic explains why.
• It is zero when nothing uses it. An NPU that no application calls saves exactly nothing, and most applications do not call it. That is measurable too, and it is the first thing to check before measuring anything else.

Three tools Windows already has

No downloads, no meters, nothing to buy.

1. The battery report. From a command prompt:

powercfg /batteryreport /output "%USERPROFILE%\Desktop\battery.html"

An HTML report with design capacity against current full-charge capacity, recent usage sessions, and estimated life. It is a history, not a live reading, so it is the wrong tool for a single task and the right one for a claim like "this update cost me an hour a day".

2. The energy trace. From an administrator prompt:

powercfg /energy /duration 60

Sixty seconds of observation, then a report naming what stopped the machine idling: processes with high processor use, timer resolution changes, devices blocking sleep. Useful for finding out what is spending your battery when the answer is not obvious.

3. The instantaneous discharge rate. This is the one that makes the experiment possible. On many laptops the battery driver reports the current drain in milliwatts, and PowerShell can read it:

Get-CimInstance -Namespace root\wmi -ClassName BatteryStatus |
    Select-Object -ExpandProperty DischargeRate

The value is milliwatts. 12000 means the machine is pulling twelve watts from the battery at that instant. Not every battery driver exposes it, and it reads zero while the charger is plugged in, which is why the procedure below insists on running unplugged.

The measurement, step by step

The task has to be something that genuinely reaches the NPU and can be turned off, which in practice means a Windows feature rather than an application you hope uses it. Windows Studio Effects during a video call is the reliable choice: it is NPU-backed on a Copilot+ machine, it runs continuously, and it has a switch.

1. Unplug. The discharge rate only exists on battery. Charge to somewhere between 80 and 60 per cent and stay in that band for both runs, because drain is not flat across a charge curve.
2. Freeze everything else. Fixed screen brightness, aeroplane mode off but nothing syncing, no browser, no updates, same power mode. Every one of those moves more watts than the thing you are measuring.
3. Baseline. Sample the discharge rate for two minutes doing nothing. Write down the average. This is what the machine costs to be switched on.
4. Run A, with the feature on. Start the call, turn the effects on, confirm the work is landing on the NPU with Task Manager's NPU column, and sample for two minutes.
5. Run B, with the feature off. Same call, effects off, sample for two minutes. For a fairer comparison run something equivalent on the CPU instead if you can, since "off" and "done elsewhere" are different questions.
6. Subtract the baseline from both. What is left is the cost of the work, and the difference between the two is what the NPU path bought you.
7. Repeat once. If the two attempts disagree by more than about ten per cent, something else was running. Find it before believing either number.

Twenty minutes, and the output is a figure in watts for your machine, not a claim about somebody else's.

Four ways to get a wrong answer

All four have caught me, and all four look like a result rather than a mistake.

The screen. A laptop panel between minimum and maximum brightness moves several watts, more than the whole effect you are hunting. If brightness changed between runs, the experiment measured the screen.

The charger. Plugged in, the discharge rate reads zero and the machine also behaves differently, because power and thermal limits are usually higher on mains. A measurement taken while charging is not a measurement.

Thermal state. A cold machine and a machine that has been working for twenty minutes have different fan speeds and different clocks. Let the fans settle before each run rather than measuring the first two minutes of one and the tenth minute of the other.

Assuming the NPU was used. The biggest one. If the work quietly fell back to the CPU, you measured CPU against CPU and found no difference, which is a true result about a different question. Check the counter first; how to see NPU activity covers exactly that, and the provider guide covers why the fallback happens without saying anything.

What to expect from the result

I am not going to print a number here and pretend it applies to your laptop, because the honest range is wide and depends on the machine, the feature and what else the system was doing. What is worth saying is the shape of the answer.

On a continuous, small, NPU-backed workload you should be able to measure a clear difference, comfortably outside the noise of repeated runs. On a one-off task lasting a few seconds you will measure nothing useful, because the fixed costs of having the machine on dominate. And on any task that never reached the NPU you will measure zero, correctly.

Whatever you find, write down the conditions with the number. A power figure without brightness, battery range, power mode and thermal state attached is not reproducible, including by you next month.

Where PC Workman fits

The discharge rate tells you what the whole machine is spending. It does not tell you where the money went. PC Workman reads processor load, package power, clocks and temperatures alongside each other, which is what turns "the laptop is drinking twelve watts" into "the processor is carrying a workload you thought was offloaded".

That pairs directly with this measurement: the discharge rate is the total, and the per-component reading is the explanation. Neither is much use without the other.

Related: is your NPU doing anything · which engine runs your model · all guides →

The commands here are standard Windows tooling: powercfg ships with Windows, and the battery discharge rate is read from the battery driver through WMI, which not every driver exposes. Platform context from Microsoft, Copilot+ PCs developer guide (NPU-backed features, and improved battery life as the stated benefit) and Windows ML execution providers (why work can quietly run somewhere other than the NPU). No figure in this guide is quoted as a result, because a power measurement belongs to the machine it was taken on.