Memtest fails off usb but not bios version

[SC724]

I recently built a new system just over a month ago and I'm trying to figure out why there's a discrepancy between running MemTest from a bootable usb flash drive, and the version built into the bios. 12900ks, Asus ROG Strix Z690-F, and Corsair 32gb DDR5 6400 (CMK32GX5M2B6400C32).

First attempt at MemTest:

[David (PassMark)]

If this is reproducible then the RAM is likely bad.

See also
https://www.memtest86.com/troubleshooting.htm

Different results on different test runs might be the result of a different test environment. e.g. ambient temperature slightly hotter. And if the memory cell is right on the knife edge of failure then there is a certain amount of true randomness that creeps in. For example maybe your voltage regulation for the RAM isn't a rock solid 1.35V. Maybe it cycles between 1.33V and V1.37 volts in a 50hz sine wave matching the mains input. Then different voltage levels would be in use from time to time when this memory cell is used. There are lots of effects like this. i.e. it's somewhat random.

Single one off failures that can't be reproduced can also be soft errors.

[SC724]

Here's the second MemTest. Same ram, just switched the slots since I was re-seating them anyway.

[SC724]

Now here's the new kit of ram failing like the original kit did.

[SC724]

Here's the same new kit, yet running off the bios version of Memtest 10.0. It's passed 3 times. But if I run it off the usb flash drive (version 10.6), it fails everytime.

Besides the MemTest version difference, I also noticed by default the 10.6 version on the usb only uses 8 cores, while the built in bios version 10.0 uses 16. Is that even a factor here? As it stands I don't trust the results when it passes with the motherboard version, but not the usb version.

[David (PassMark)]

Strange that two kits would fail in almost the same way. Both kits seem to be have identical model numbers. Are you sure are that last screen shot was a new kit?
Do you have the final HTML report from each run as that should include the serial numbers?

It does look like bad RAM however. Very hard to explain reproducible single bit errors at a single address otherwise.

While this doesn't exactly match your issue, there was also the recent case of systematic failure of Corsair RAM. Which might point to general manufacturing / QA issues.

Update: The 8 v 16 cores thing is because we stopped using hyperthreading cores in more recent releases. Using hyperthreading (paradoxically) made the testing slower, so there was less load on the RAM.

[SC724]

I have pictures of both kits including serial numbers. Nothing is overlocked. XMP 1 is used in bios, however it failed on XMP 2 as well. Plus the ram is on the QV list of this motherboard. I noticed 10.0 that passed didn't show the entire model number. Thought that it was just because of an older version of MemTest.

[SC724]

The 12900ks is a 16 core cpu with 24 threads. Not an 8 core cpu.


How can both sets of ram be bad when they pass on bios MemTest, but not usb version? Shouldn't they pass/fail on both?

[David (PassMark)]

I think the errors are a real hardware failure of some sort.
If you have a memory kit from other vendor, I would try that next.

[SC724]

I'm ready to replace the Z690 with a Z790 and/or even go back to MSI. You don't seem to suspect a motherboard issue. I'm at a loss but I really appreciate all your expertise in this matter.

I have read that 6400 is the limit of Z690 boards. But again, it's on the QV list. I could try a different vendor as you suggested, but I also don't trust the motherboard now either.

Everything is returnable until the end of this month.

[David (PassMark)]

It is easy to turn off XMP as an experiment and run the RAM at lower JEDEC default speeds.

[SC724]

What would default speeds tell you? Couldn't the motherboard still be at fault here? Why do you suspect it's a ram issue and not a motherboard issue?

[David (PassMark)]

It is very hard to explain reproducible single bit errors at a single memory address if the problem is with the CPU or the motherboard.

A lot of RAM errors aren't 100% bad cells (meaning that they fail all the time in all circumstances).
They are more commonly weak cells. The 100% bad cells are picked up by even the most superficial testing during manufacturing, so they aren't so common in the field, at least in new RAM.
Many RAM errors are thus related to external factors that trigger the weakness. i.e. they only fail some of the time, or only over certain speeds, or only over certain temperatures, or only at certain voltages, or they are venerable to EMI via row hammer, etc...

So if you had weak RAM, it might come good at slower speeds or higher voltages.

[SC724]

Ok. Once again I appreciate your help and knowledge. Also every time it's failed was either test 5 or 7. I'm going to order a new kit from a different vendor and try that next as I have no other DDR5 ram here.

Update: Ordered a new kit from GSkill. Should be here tomorrow. I'll post back with my results. Thank you!

[SC724]

Great news! Recieved the new GSkill 32gb DDR5 6400 (F5-6400J3239G16GX2-RS5K) and it passed for the FIRST time in MemTest 10.6 from usb flash drive as well as the bios version 10.0. I've never had two defective ram kits back to back. Corsair really needs to address this. I've used both GSkill and Corsair for various builds with zero issues over the years. Asus ROG Strix Z690-F with latest bios 3101 (XMP 1).

Thank you so much David for everything!


MemTest 10.6 off usb flash drive, second is MemTest 10.0 via Asus bios