# GetMOBILE article on benchmarking micro-NPUs now out

*2026-08-05 — note*


<a href="https://dl.acm.org/doi/pdf/10.1145/3833428.3833430"> <figure class="image-right-float"><img src="/images/npus-ss-article-1.webp" alt="Benchmarking low-power NPUs at GETMOBILE" title="Benchmarking low-power NPUs at GETMOBILE" loading="lazy" srcset="/images/npus-ss-article-1.768.webp 768w, /images/npus-ss-article-1.640.webp 640w, /images/npus-ss-article-1.480.webp 480w, /images/npus-ss-article-1.320.webp 320w, /images/npus-ss-article-1.1600.webp 1600w, /images/npus-ss-article-1.1440.webp 1440w, /images/npus-ss-article-1.1280.webp 1280w, /images/npus-ss-article-1.1024.webp 1024w"><figcaption>Benchmarking low-power NPUs at GETMOBILE</figcaption></figure> </a>
[Josh Millar](https://profiles.imperial.ac.uk/joshua.millar22) lead an article just out in [GETMOBILE](https://www.sigmobile.org/pubs/getmobile/) on our work from last year on [benchmarking micro-NPUs](https://anil.recoil.org/papers/2025-npu-bench) to help make sense of the growing number of dedicated, ultra-low-power 'neural network accelerators'.
There are lots of hidden embedded NPUs in modern low-cost chips, ranging from [ESP32](https://www.espressif.com/en/products/socs/esp32)\-based boards over to [ARM](https://arm.com) ones. All of these have quite a variety of tradeoffs, from the operations supported (which affects which models can be run on them) to the amount of memory and CPU power.

If you'd like a shorter, more accessible summary of that paper, then [head over](https://dl.acm.org/doi/pdf/10.1145/3833428.3833430) to read our article.  Feedback is most welcome\!

While the hardware landscape hasn't changed hugely since last year's article, the software one has. I'm now finding that using coding agents to write low-level embedded code is far far faster than before. Modern harnesses are pretty good at figuring out the various quirks of each cross-compilation and embedded environment...
Synopsis: Resource-constrained, MCU-scale neural processing units have emerged to provide real-time inference within milliwatt-scale power budgets.
Words: 161

## Related

- [Benchmarking Ultra-Low-Power -NPUs](https://anil.recoil.org/papers/2026-benchmarking-npus) (paper, 2026-08-01)
- [Benchmarking Ultra-Low-Power μNPUs](https://anil.recoil.org/papers/2025-npu-bench) (paper, 2025-11-01)

---
Canonical: https://anil.recoil.org/notes/sigmobile-npu
Type: note
Tags: esp32, embedded, sensing