Ironically, the NPU was supposed to do the AI work instead of the GPU, but instead it's used to work as the GPU to unload AI work from the GPU.
GPU and NPU are two parts designed to do a lot of parallel math, the first at the highest performance achievable, and the latter at the lowest power budget possible.
It’s always good to put idling parts to some use. Some AI workloads can probably run well on the DSPs.