If they thought that peak performance didn't matter, they wouldn't quote peak performance numbers in their comparison, and yet they did. Peak performance clearly matters, even in battery powered devices: many workloads are bursty and latency matters then, and there are workloads where you can be expected to be plugged in. In fact, one such workload is generative AI which is often characterized by burst usage where latency matters a lot, which is exactly what these NPUs are marketed towards.