Beyond the 29 TOPS rating, do the 245's 14 cores actually stay busy in desktop AI work?

  They stay busy once the division of labor is clear. The NPU owns always-on loads that should not burn watts, while the 14 cores with VNNI step in for short quantized small-model runs at a 5.1GHz boost — the practical ceiling there is dual-channel DDR5 bandwidth, not thread count, and a desktop's generous cooling lets such bursts run free.

  An office desktop that wants light AI resident plus the occasional local model inference fits the 245 neatly. Expecting near-GPU generation speed from it is a category error. Sort out first whether this machine's AI time is mostly background or mostly compute: it excels at the former, and the latter belongs to a discrete card.