The Procyon AI benchmark measures inference times on popular models such as MobileNet, ResNet, and YOLO. It's designed to show how well a system can accelerate machine-learning tasks in real use.
Procyon scored 121 in its CPU-based computer vision test, with MobileNet V3 averaging 1.04 ms and the heavier REAL-ESRGAN model taking 4,317.19 ms. These are CPU inference results using Windows ML, not measurements of a dedicated NPU, and I'd view local AI work as an occasional extra, not a reason to buy the X3.