Being quite energy efficient and quite fast especially regarding the time to first token of responses, one use case I've envisioned so far for npu is almost-realtime-voice-transcription without boggling the cpu/gpu hw. My prototyped experiments into this field so far beared to early nice results mostly by accessing the fastfloflm server (part of the lemonade framework yet independently usable) serving whisper models. Hopefully more will come if/when the iron compiler will be incorporated not just into windows only copilot+ stuff but instead also into projects like llama.cpp (or maybe into vllm) and also when it will be more feasible to indipendently train/quantize/fine tune more recent models for running into npu and in hybrid mode (that should be sort of speculative decoding, where prompt/rag is first computed by the fast but less capable npu and then passed to the more powerful but slower gpu) than the one shared by amd through onnxmodelzoo and similar repositories. Other exciting use cases could be related fast visual recognition

on a side note, I've just discovered that on strix halo, the npu power mode could be set from performance (default) to turbo mode through the command xrt-smi configure -d 0000:c6:00.1 --pmode turbo. Still to be tested tho