I think of Aida64 as a functional exercise. It can run through some functions to see that the rig is basically working and to check cooling without much risk. However it is neither stressful nor a test -- the results are not checked. Only the memory bandwidth benchmarks are accepted as meaningful tests.
I'm also down on Prime95 for the reason that it does not test -- doesn't check the math results. After all that stress, the CPU and RAM can still be unstable. I have experienced the degrading that Arne mentions after running a CPU at high stress. It's real and Prime95 is known to be hard on CPUs.
My favorite is y-cruncher (look it up on HWBOT or google it). It's highly effective for all the above areas except GPU and video RAM. It catches the slightest bit error in memory or computation. Memory use is intense enough to kick up errors that are not shown by anything else but HCIMemtest. CPU processing uses all the computation capability of the CPU, including AVX - it picks different code modules for different generations back to Sandybridge plus a generic legacy version up to AVX-512 in latest generation XEs. As such it runs the CPU hot for a satisfying stress test and good cooling is essential.
Test time is proportional to the number of threads. I used it for memory tuning on the M12A last week - 45 seconds with 16 threads for the 1 billionth digit of Pi. That scales to a minute and a half for 8 threads in a 4-core CPU. Compare one minute typical test time that covers CPU and memory stability to hours for Prime95 plus HCIMemtest.
When using y-Cruncher for memory tuning, be sure to have a fully stable CPU profile and vice versa -- be sure to have a stable memory profile when testing CPU overclock. The test behaves the same when errors from either component are caught. Isolating the cause depends on trying one thing at a time so you know what to undo if/when the overclock or memory tweak fails. One memory tweak at a time isn't so tedious when the test only takes a minute.