ITdaily ran local AI on 15,000+ photos for several hours. Faces were usable after merges; object tags misfired. No watt-hour savings figure and no ARM VM test.
Test setup, measurements & limitations
Qnap TS-AI642 review: 6 bays, plenty of power and a (sometimes) useful NPU
Tested: QNAP TS-AI642 with 8 GB RAM; QTS 5.2; just over 15,000 mixed photos for local NPU face/object/similar-photo analysis. No named HDD RAID or Ethernet throughput tool.
Applies to: Exact 8 GB TS-AI642 memory and NPU photo features. Not a sequential file-copy measurement. Networking description (2x 1GbE + 1x 2.5GbE) matches the official table.
- Just over 15,000 mixed photos took several hours in background NPU jobs. Face recognition usable after merging groups; object categories often wrong. No power meter and no non-NPU A/B, so no efficiency percentage.Source location: Faster and more efficient / Cat in a bag
- Do not copy generic Virtualization Station marketing onto this ARM SKU as a tested VM result.Source location: QTS 5.2 paragraph
Measurements reported by the reviewer| Measurement | Result | Conditions & source location |
|---|
| Photo library processed | >15,000 photos, several hours, background NPU | Mixed devices/years; not a MB/s disk test.Faster and more efficient |
|---|
Limits to keep in mind
- No MB/s RAID test.
- No instrumented NPU vs CPU power comparison.
- Euro street price is not a US offer.