Nothing you’ve said contradicts anything I’ve said, except for a few minor errors you’ve made. I still don’t know what you disagree with me about.
Your terminology is all around the place, and you keep arguing about insignificant things while ignoring the important ones. Yet you say I am the one who is making errors.
You keep telling me things I already know, implying you’re disagreeing with me about something.
I am disagreeing with claims that Dynamic Caching has significant impact on system RAM usage. Dynamic Caching is about managing of internal GPU resources.
The whole logic is probably done in the memory controller.
This doesn't make any sense. Memory controller does not even come into play at the level or GPU register files.
BTW, there is no system RAM. When someone says system RAM, that is defined as the standard memory available to the CPU, also implying there is dedicated graphics memory, which there isn’t. In the case of the M-series of chips, the unified memory is used by both the CPU and GPU.
System RAM = the DRAM in the system. As opposed to private memory buffers and caches that live on specific processors (such as the fast on-core buffers used by the GPUs to accelerate local processing). I have no idea why you think that the term implies existence of separate graphics memory pool, but ok.
The GPU (or CPU) makes RAM requests to the MMU and the MMU allocates and deallocates accordingly.
The MMU does not allocate or deallocate anything. Its job is to translate addresses. Allocation of system RAM is the job of the OS kernel. In the context of the Apple patent, the private on-core MMU might have extended functionality (such as reserving local scratchpad memory if the simdgroup requests it), but that has nothing to do with the main system MMU.