Looks like their NPU (aka ANE) takes up about 1/3 of the die area of the GPU.

Would be interesting to see how much they’re _actually_ utilizing the NPU versus their GPU for AI workloads.