Thanks, I will investigate and profile. Perhaps you know why the Compute timeline shows a maximum of 2 lines with running Command Buffers? Could it be a design or is there another reason?

Yes, I tried both MTL::DispatchType::DispatchTypeSerial and MTL::DispatchType::DispatchTypeConcurrent. does not matter, behavior does not change. I use MTL::DispatchType::DispatchTypeConcurrent.

Computing do not depend on each other. About MTLHazardTrackingMode: Yes, by default I set MTLResourceStorageModeShared, changed to MTLResourceStorageModeShared | MTLResourceHazardTrackingModeUntracked and really saw the parallelism in the profiler (see picture). But the performance has not changed and with 1 MTLCommandBuffer in sum is better than when I create 2 or more. Also, as you can see from the figure, it does not matter what kind of parallelism I create, in the profiler it draws only two parallel streams (timeline Compute).