What is an NPU?
A neural processing unit is a small accelerator built into a processor for AI tasks. It runs certain models efficiently at low power, which suits a NAS that is always on. Its speed is quoted in TOPS, trillions of operations per second.
Which NAS systems have one?
The table on this page lists every system we track with an NPU, from the live records. The main groups are:
- Intel Core Ultra systems such as UGREEN's iDX6011 and iDX6011 Pro and Zettlab's Ultra models, with NPUs of around 11 to 13 TOPS
- AMD Ryzen AI and Ryzen PRO systems: the Minisforum N5 Pro with 50 TOPS and the NIMO Nexus Pro with 16 TOPS
- Rockchip RK3588 systems such as the Zettlab D6 and StationPC DA400, with 6 TOPS
Some vendors quote a platform total that adds CPU and graphics to the NPU; we record the NPU alone.
What does a NAS NPU actually do?
It runs the vendor's own AI features: face and object recognition in photos, scene detection, transcription, smart search and camera detection. These use models the vendor has built for that specific chip.
How do you know whether a NAS has an NPU?
Check the processor model rather than the marketing. Intel Core Ultra and AMD Ryzen AI chips include one; AMD Ryzen 8000-series PRO chips such as the 8845HS do too. Older Ryzen V-series and R-series embedded chips used in many Synology and ASUSTOR models do not, and neither do Intel Celeron, Pentium and Core i3-N parts. Each system page shows the NPU figure from the manufacturer where one exists.
Does an NPU make local LLMs faster?
Not with today's common tools. Ollama and llama.cpp run models on the CPU or a GPU, and NAS operating systems do not yet expose the NPU to them. Minisforum documents running models on the N5 Pro's NPU, and StationPC's DA400 uses a separate AI module for small models, but these are vendor-specific routes.
For language models, memory and a GPU path matter far more than NPU TOPS.
Common questions
Is more TOPS always better?
Only for the tasks the vendor's software runs on the NPU. Compare memory and GPU options first if local LLMs are your aim.
Does the QNAP TS-233 have an NPU?
Yes, a small one used to speed up face and object recognition. QNAP does not publish a TOPS figure.
Can I use the NPU with Ollama?
Not today. Ollama and llama.cpp run on the processor and on supported graphics, and do not use the NPUs in current NAS processors. Some desktop apps use NPUs, but those are not what a NAS runs.
Which NAS processors include an NPU?
AMD Ryzen AI processors, such as the Ryzen AI 9 HX 370 and the Ryzen AI Max range, and Intel Core Ultra processors include one. The Intel N-series chips found in many budget NAS systems do not.
Will NAS makers add more NPU support?
Possibly. Software support for NPUs is improving, and NAS makers are adding features such as photo recognition that can use them. Treat the NPU as a bonus, and buy for memory and GPU options if you plan to run language models.
Is an NPU useful for photo recognition?
It can be, where the NAS software uses it. Face and object recognition runs in the background across large libraries, so moving it to an efficient NPU keeps the processor free. Check whether the maker's photo app uses the NPU before counting on it.
