But then why do my apps not use more than 16GB RAM? Everything together never surpasses 16GB RAM (maybe a bit, because it gets compressed). And in the meantime the GPUs show <1% RAM usage.
And Ubuntu has no issues with it, while Windows crawls on its knees once that limit is reached.
Often that reserved memory is a soft limit that programs try to respect, but they can force allocate more. So what happens when you disable page file, close everything, open python3, and type this?
array20gb = bytearray(20 * 1024**3)
This should allocate 20GB of RAM to that process, starting with unused unreserved RAM, and then it should take your shared video memory after the unused unreserved RAM is allocated. If you get an error, then you know that reserved video RAM can't be used by programs.
Thanks! This is odd, and indeed it does start filling the RAM.
What could it then be what hinders the allocation of more than 50% under "normal operation". Initially it's fast, but over the course of a couple of hours if gets so low, that it takes around a minute to open a folder from the desktop via double-click while memory is at 50%. Temps are OK at that time and nothing is really getting used, like SSD is at 1-5%.
Probably system or driver issue resulting in lending/grabbing the maximum amount for no reason. Performance-wise, 4 GB dedicated video card gets no benefit from those extra gigabytes in RAM buffer anyway.
Do Quadro cards support some kind of “fat and slow” compatibility mode for software that demands giant memory size on a professional card to fit all the data there, and can wait to get the results? Maybe there's some switch you've toggled to enable “20 GB VRAM” mode?
And Ubuntu has no issues with it, while Windows crawls on its knees once that limit is reached.