PC Memory Timing Strangeness

[Coeus]

I recently upgraded my main PC at home from 16Gb to 32Gb by adding two new 8Gb modules. Possibly related, the PC has occasionally frozen since. I can't be sure the two are related as there have also been kernel upgrades in the same period of time.

I have noticed something a little strange, though. The memory that was in the machine previously has this sticker on it:

The additional memory is described as having the timings 9-9-9-24 so slightly faster, if I have read the info correctly.
When I went into the BIOS advanced mode to see what memory timings it was using it was 9-9-9-24. So I am beginning to suspect this may be the cause of the trouble, i.e. it is essentially "overclocking" the older memory.

I gather the memory modules have a serial EPROM with the timings, and maybe other things too, which the BIOS should be reading. If so it looks like it is being lazy and only reading one module and not doing what I would expect which is to check all the modules and then use the timings for the slowest module found.

Anyway I have now manually set the timings to the slower set and I'll see if it makes any difference.

Interestestingly, MemTest86 didn't find anything.

[richardtoohey]

Will be interesting to see if the change works.

If it's not DNS, then it's usually RAM! Not sure what my third rule should be (mind you, usual-problem 1 should be the end-user, myself included!)

[SarahWalker]

Third rule would be PSU in my experience!

[Boydie]

May be as well to check the voltage supplied to the memory as well. If your bios is selecting that from the new memory’s profile as well, and it’s only asking for 1.35V then the old stuff won’t be happy...

[Coeus]

May be as well to check the voltage supplied to the memory as well. If your bios is selecting that from the new memory’s profile as well, and it’s only asking for 1.35V then the old stuff won’t be happy...

Thanks for that. I've checked and both are 1.5V. That's good as the BIOS doesn't seem to allow me to change that anyway.

[Boydie]

Try 24hrs Prime95 on Blend?

Since you don’t know which (if any) component is flakey, Blend will test all of them. If it throws up errors, the culprit FFT size will help track down if it’s memory/cpu etc...
Then try individual sticks if it looks like memory.

How long did you run memtest for (and I presume it’s a recent version - older ones ran fewer tests)?

Although you shouldn’t need to change them for stock memory speeds, are your vccsa and vccio sensible (if your bios lets you alter them)? Could adding more ram have exposed a dodgy IMC?

I really hate memory issues. One of my recently retired motherboards had a slightly dodgy connection to one of the channels - worked fine with both sticks in one channel, but for dual channel if the cpu heatsink was screwed down *slightly* too tight the miniscule flexing of the motherboard was enough to make it fail.

[Coeus]

Try 24hrs Prime95 on Blend?

I have not heard of that - I'll look it up.

Since you don’t know which (if any) component is flakey, Blend will test all of them....

I am not even sure the problem is hardware. The reason for suspecting the memory is simply that I had recently added some and things that have changed are usually more likely to be the cause of a problem than things that have stayed the same.

How long did you run memtest for (and I presume it’s a recent version - older ones ran fewer tests)?

The first time for just the one pass. I left it overnight last night to run four passes (the most the proprietary but free of charge version will do) and it found no errors:

Although you shouldn’t need to change them for stock memory speeds, are your vccsa and vccio sensible (if your bios lets you alter them)?

I don't think it does - I can't find anywhere to change any voltages.

Could adding more ram have exposed a dodgy IMC?

Could you explain, please?

I really hate memory issues. One of my recently retired motherboards had a slightly dodgy connection to one of the channels - worked fine with both sticks in one channel, but for dual channel if the cpu heatsink was screwed down *slightly* too tight the miniscule flexing of the motherboard was enough to make it fail.

I have definitely had connection issues in the past but these have usually shown up pretty quickly and have always been resolved by reseating the memory stick in the slot.

[Boydie]

The IMC question was speculative on my part - if the cpu has a slightly weak/elderly integrated memory controller it may be okay for two sticks but unstable with four. Checking the forums out, this can indeed be an issue, more so if the ram is being overclocked, but also at stock.

What is your cpu? How old is it?

If Prime95 does throw something up (the ram may be stable when that’s all that’s being tested but may be iffy when the whole system’s under stress), it may be worthwhile checking what voltages your motherboard is providing - HWInfo64 is very informative for these.

From personal experience, I’ve found stock motherboard settings to be way too high - whilst I overclock some of my systems, they’re pretty much all undervolted relative to motherboard auto settings (worst offenders are Skylake systems - I undervolt my laptop by -0.155V and the desktop by -0.05 despite overclocking all cores to 4.3).
Unnecessarily high voltages can increase heat and voltage degradation over the years, so loony auto settings may cause failure over time even if everything is stock. Plus everything runs cooler and quieter, with less thermal/tdp throttling if you undervolt (esp on the laptop).

[Boydie]

Coincidentally, I’ve been suffering random restarts on my latest project.
The DDR4-3200 was overclocking to 3600 (with relaxed timings), and passing Memtest86 absolutely fine, so that couldn’t possibly be the issue, right?
The rest of the system hadn’t been altered, but Intel BurnTool was no longer stable. Adding vcore couldn’t stabilise it.
Dropped the ram back to stock speeds, and the problem went away.

So, today I learnt that Memtest pass doesn’t prove the memory isn’t the issue.
I’m just making sure everything’s fully stable, then I’ll try again, only with higher System Agent, to see if that can stabilise the memory overclock.

[Coeus]

So, today I learnt that Memtest pass doesn’t prove the memory isn’t the issue.

That's interesting. It does seem to be do quite a comprehensive set of tests and can any set ever be exhaustive?

In the meantime things do seem to have improved in that the only recent lock-ups have not been completely random and only happened trying to start b-em and only after the machine had been suspended. They never happened when the machine was freshly booted.

Thinking this may be a s/w issue I tried various older versions of the kernel, mesa and xorg-server but nothing made any difference. Then I remember I had configured hybrid-sleep so it writes the memory to disc and then sleeps so that if the power should fail while sleeping it would just reload the memory from disc. Changing that back to normal (in RAM) sleep solved that. That does seem like a bug, though.

[BigEd]

I'm not sure how good memtest is, in practice. I've seen it spot errors but more often fail to find a problem when I rather suspect there is one.

How about trying y-cruncher? From the FAQ:

y-cruncher has a reputation of being notoriously stressful on the memory subsystem. If you read around overclocking forums, there are countless reports of y-cruncher being able to uncover subtle memory instabilities where all other applications and stress-tests will pass.
It's worth noting that y-cruncher was never designed to be a memory stress-test. And to date, it is still unclear why it often seems to be better at testing memory than even dedicated tests such as MemTest64.

[Ramtop]

That's interesting. It does seem to be do quite a comprehensive set of tests and can any set ever be exhaustive?

The problem with Memtest is its generic nature as it's written to run on such a wide range of hardware. A really punishing memory test would need to be aware of the specific CPU and entire memory sub-system.

As is stands Memtest is fine for catching egregious memory errors like failed bits, but anything more elusive and it isn't worth much. I've worked on PCs that will run Memtest fine for days but can't even boot Windows because of memory issues.