Got a frame capture. Unfortunately MTLReplay can't deal with traces this big in full profiling mode, but still, got a frame capture from the benchmark
View attachment 34328
I'll start looking at and doing whatever analysis I can - but I have very little Metal experience so it may be limited, but here's the gputrace file if others want a peak (you can of course also take a trace yourself - In Xcode, select. debugging the app, set the environment variable "MTL_CAPTURE_ENABLED=1", enable metal frame capture in options, select the M in the debug bar and set a shortcut like cmd+F1, run the benchmark and hit your shortcut - then wait an absolute eternity. At some point Xcode will trick you into thinking there's a problem by saying "No GPU Workload Detected"; Ignore and let it continue to run - eventually it'll beachball. Leave it for a while, then the MTLReplay pops up. Game remains beach balled forever so just kill the process when you're done with it.
Once again the archive is way too big for the website so download is here:
I will say this though, the trace looks extremely odd to the point that I don't really trust it. The thumbnail for the present frame looks wonky and the performance counters all suggest 0% occupancy on all blocks. Occupancy may be because it's in Lite mode, but you'd think the render steps could still be pictured correctly