I tested your code with the latest version:
tensorflow-mac==2.7.0
tensorflow-metal==0.3.0
The bug still is there:
CPU Prediction: 4.890502452850342
GPU Prediction: 3.1839447021484375
Indeed, very concerning!
Hopefully someone from apple sees this!!
Ps: other reports:
https://developer.apple.com/forums/thread/693178
https://developer.apple.com/forums/thread/695256 (the author of this post (@radagast) has multiple posts on this issue)