Basic repro case based on CPU-GPU-Synchronization sample does not have this problem "unfortanutely". This is probably a combination of multiple factors that cannot be that easily recreated in a sample project because, well, our code base is big and lots of Metal features are used for every frame it renders. I'm afraid I can't even share binary right now :(

Hi. Thank you for the response. Fast math is disabled already :( I'll try to create a sample project and post feedback.