It's quite common "thermal condition" for any device running Augmented Reality app. ARKit, RealityKit, ARCore, Vuforia or MRTK's tracking stage is highly CPU-intensive. Your phone not only tracks and reconstructs a surrounding environment at 60 fps but also simultaneously renders 3D geometry with PBR shaders, textures, shadows, animation and physics.

In some cases, Face tracking is even more CPU-intensive than World tracking. This can be possible due to the point that RGB channels coming from the selfie camera are in tandem with a segmented Alpha channel and ZDepth channel, coming from TrueDepth sensor. And there are more than 50 facial blendshapes deforming geometry at 1/60 fraction of a second.

Pay particular attention to the fact that native Xcode builds of ARKit apps written in Swift (using UIKit or, especially, SwiftUI) run considerably faster than Unity builds of ARKit apps.