I've been playing around with it on M1 Max since the Alpha build with great results.
Here's some numbers from my latest benchmark run:
"Junk Shop" demo using Cycles
M1 Max (10 cores CPU / 32 cores GPU)
Blender 3.0 (CPU) / 1:14.64 minutes
Blender 3.1 (CPU) / 1:10.86 minutes (1.05x)
Blender 3.1 (Metal GPU) / 0:44.16 minutes (1.68x)
Blender 3.1 (Metal GPU+CPU) / 0:27.90 minutes (2.67x)
For a reference, RTX 3090 (OptiX) was 0:18.22