The marketing figure of 45 TOPS is a hardware ceiling. The question is which workload approaches it.
The test rig
Three devices, same build, same model. Tests: (1) local language model inference, (2) image generation, (3) video background blur, (4) meeting transcription. Three runs each, cold start included.
Results
On the language model the NPU runs 2.4x faster than CPU — but only with a quantised model. At full precision the gap is 8%, inside measurement error. Image generation barely touches the NPU: the driver still routes to the GPU.
Battery: the real win
Six hours forty minutes of video conferencing, against four hours fifty-five on the same device without the NPU. That is real and measurable. The true part of the advert is duration, not speed.
Buying advice
Do not buy a device for the NPU today. Buy it for battery life and quiet operation. The picture changes when second-generation software support lands.
Comments
Related stories
The quiet switch era in mechanical keyboards
New springs sit around 45 grams and suit office use. But what will players say?
Room correction: do not set EQ before you measure
Eight speakers, two rooms, four mic positions. Seventy percent of problems below 200 Hz come from one corner.