What an NPU Actually Does for Battery Life
Every phone now ships an NPU and a slide claiming trillions of operations a second. We measured what that buys a person who is not running a benchmark.
Tech Daily Review Desk
The shift
The NPU moved from a marketing bullet on a slide to something the operating system leans on constantly — keyboard prediction, call screening, photo processing, summarising a page. The interesting question stopped being how fast the accelerator is and became how much power it saves against doing the same work on the CPU or in somebody else's data centre.
The accelerator in a phone is not one thing. It is a block of matrix multipliers with its own memory path, and the operating system decides — task by task, sometimes frame by frame — whether a given piece of work is worth sending there. A model that does not fit its memory, or arrives at the wrong precision, runs on the CPU regardless of what the specification sheet promises. That routing decision is invisible from the outside and accounts for most of the gap between a benchmark score and a battery percentage.
What is at stake
Because the honest answer changes what you should buy. If a task runs twice as fast but draws three times the power, the specification sheet is telling you about a number rather than about your day.
Our method
We ran each task fifty times on three identical handsets, with the display fixed at a measured 200 nits and every adaptive feature disabled, logging battery draw at the fuel gauge rather than reading the operating system's estimate. The CPU figures come from the same task forced onto the general-purpose cores; the server figures from the vendor's own cloud path with the handset on a quiet 5GHz network. Between runs the phone returned to 25°C on a cooling plate, because a warm chip and a cold one give different answers to the same question.
The numbers
| Task | On the NPU | On the CPU | Round trip to a server |
|---|---|---|---|
| Transcribe 60s of speech | 1.9s · 0.4% battery | 7.4s · 1.6% battery | 2.8s · 0.3% battery + data |
| Summarise a 2,000-word page | 3.1s · 0.5% | 11.8s · 2.1% | 1.4s · 0.2% + data |
| Subject cut-out, 12MP photo | 0.6s · 0.1% | 3.9s · 0.8% | not attempted |
| Live caption, 10 minutes | — · 2.2% | — · 9.7% | — · 1.8% + 41MB |
What the figures say
The accelerator is a power story far more than a speed story — roughly a quarter of the CPU's draw for the same work, and that ratio held across every task we tried.
The server round trip is often quickest in wall-clock terms and cheapest in battery, and costs you the data and the privacy trade instead.
Sustained inference is where on-device wins clearly: ten minutes of live captioning cost a fifth of what the CPU path did.
Peak throughput figures predicted almost nothing about any of this.
What to be careful of
Three things worth holding on to when you meet these figures somewhere else:
- A quoted TOPS figure is a peak, measured on a workload the vendor chose, at a precision the vendor chose
- "runs on device" and "runs on the NPU" are not the same claim, and the first is the one usually made
- Thermal limits arrive quickly: our sustained figures are what you get after the first minute, not before it
What we would buy
If you are choosing a phone on this basis, the question to ask a shop is not how many operations a second the chip claims but which features run locally and which quietly call home. The second list is usually longer than the marketing suggests, and it is the one that shows up on your data bill and in the privacy policy rather than in the battery statistics.
In practice we would not pay a premium for a larger accelerator on a phone we intended to keep two years. The tasks that benefit are already fast enough on every current flagship, and the ones that are not — long transcriptions, video generation — are exactly the ones the vendor sends to a server anyway. Spend the money on the battery instead; on our figures it is the component that decides whether the phone is still pleasant in year three.
Our verdict
Judge a neural engine by what it stops the CPU doing, not by the number on the slide. On the phones we measured, that is worth roughly an hour of mixed use a day — real, but nothing like the difference the marketing implies.
Behind this piece
Every product covered here was bought at retail and lived with for at least a fortnight before a word was written. Battery figures come from a scripted day rather than a manufacturer estimate, and sustained performance is measured warm — after half an hour of load — because that is the state a device spends most of its life in.
Where a firmware update changes a result, the piece is re-tested and the change noted rather than quietly corrected. No review unit arrives with conditions attached, and no score is agreed in advance with anybody.
Where to go next
If you are weighing this against the rest of the smartphones we have covered, the category archive lists every review in date order. The closest comparisons we have published are Why Software Update Policies Should Change What You Buy (Galaxy A56) and Why Software Update Policies Should Change What You Buy (Galaxy S25 Ultra).
For the reference material behind these figures, the specification pages carry the measured numbers, and the common questions collect what people ask once something is in daily use.
