memtest86 4 pass run times - slower on 6 cores than 4?

[rjwilmsi]

TLDR: is memtest slower on a 6-core CPU than 4-core i.e. does it scale negatively beyond 4 cores?

I've been investigating the run time of memtest86 on different hardware.

Using the same 2x 16 DDR4 2666 MHz UDIMMs (JEDEC, no XMP profiles), I get the following results for the 4 full passes in the Free version:

CPU	Mobo	Speed	Memtest version	Run time
Intel i5-6400 (4c)	ASUS H110M-R	2133 MHz	9.3	4h 17m
AMD Ryzen 5 3600 (6c)	Asrock AB350M Pro4	2666 MHz	9.3	6h 40m
Intel i5-8500 (6c)	HP 400 G5 MT	2666 MHz	10.0	5h 55m

(Note: Intel i5-6400 only supports 2133 MHz at JEDEC speeds)

The result I would specifically like comments on is the i5-6400 versus i5-8500: per Intel the i5-8500 has higher base clock and higher turbo clock so surely memtest86 should not be slower (it may not be much faster if limited by memory bandwidth but surely should not be slower). Secondly the RAM is running at 2666 MHz rather than 2133 MHz (testing with DDR3 RAM on another system I found that as expected faster RAM did lead to slightly faster run times while keeping RAM size, CPU, motherboard the same). I monitored the runs and CPU temperatures never went over ~70C so I don't think any thermal throttling occurred.

So is the implication that using more than 4 cores actually makes memtest86 run more slowly? If there's some published info on run times / core scaling please let me know, I could not find it. Is there a way I can test that on the Free version or do I need Pro to run in parallel but on a reduced number of cores?

[David (PassMark)]

If you are doing benchmarking, then you need to keep the test conditions the same and vary just a single factor at a time.
e.g.
Same hardware but different version of MemTest86.
OR
Same version of MemTest86 but different hardware.

But yes, we did observe lower performance with a high number of cores. The explanation is pretty simple. More CPU cores doesn't make the RAM any faster. But does make the cache less effective and adds task switching overheads. The point of diminishing returns depends of the relative RAM & CPU performance. Number of memory buses, dual or quad channel mode, size and effectiveness of the CPU cache, etc... There isn't a simple answer to the optimal number of cores.

This also also why the Free edition of the software only supports 16 cores and why we disable hyper-threading.

Sometimes BIOS bugs also stop MemTest86 from running in more than one thread.
There are also cases for Test 13 where if MemTest86 suspects an error, it might do an additional longer run to verify there really is a fault. Would really need to look at the log from each test run to verify this. But that 6h40min run does seem a bit slow.