What is the optimal number of CPU cores for testing RAM with MemTest86

[David (PassMark)]

While developing V12 of Memtest86 we were looking at a couple of the slower longer tests and if they could be made faster.

There was some minor improvements that could be made for a single threaded test (where the memory bandwidth wasn't saturated, because a single core isn't fast enough), but for a multi-threaded test the actual test loops were already pretty good for most hardware and they saturated the available memory bandwidth.

Here are the results with the V12 optimized test Conclusion: You don't gain much testing speed after 4 threads. The same would generally hold for V11 as well.
(Note: on systems with slow CPU cores and very fast multi-channel RAM, scaling would likely continue. Maybe up to 8 or 10 cores).

Only 1.4x faster going from 1 to 6 cores for Test 9. And 2.8x for Test 7.

Tested on Ryzen 5 5600X with DDR4 3600MT/s RAM.

[David (PassMark)]

And as a tease of V12 here are some numbers for V11 vs future release V12 code for these two tests.
Almost double the speed from Test #9 in V12, but only for a single thread, without changing how the test works. But by the time we get to 6 or more threads they are all waiting and idle as the RAM isn't fast enough. RAM becomes the bottleneck.

In short: code efficiency matters a lot for single thread (or very slow CPUs), but doesn't matter at high core counts, or with slow RAM. Having many CPU cores can hide inefficient code.