90s apps used these Windows handle controls and it was not a performance problem. On computers 100 slower than today and with 1000 times less memory.
BTW, the current Windows Task Manager still shows the total HANDLE count on the CPU page.
They were not used in apps with complex UIs because native controls were hard to skin and didn't work very well when their size was very small (as complex UIs tend to need)
Tensor cores massively accelerate matrix multiplication with no algorithmic breakthrough. Just by being smart about how you move/cache/reuse values, operations which are considered O(1).
There should be an O notation which takes into account everything - various kinds of memory latencies, speed of logic operations, .... Obviously we have wall clock or total energy used.