TridentZ 3000C14, memtest vs aida64

[mackerel]

Earlier today I receive G.skill Trident-Z 3000 C14, which I bought to try some B-die OC. My Ryzen 1700 system was the nearest free one to try it on. I swapped the modules in, loaded XMP, sorry, DOCP, and let it go at 2933 C14 in memtest. This competed a pass without error. Feeling lucky, I went to 3200 15-15-15-36 and it booted to Windows. Still feeling lucky, I dropped it to 3200 14-14-14-34, and went straight to memtest. That completed more than one pass without error. I rebooted into Windows, started Aida64 (latest non-beta). If I run a stress test selecting only cache and ram, it showed an error almost instantly. I deselected cache, started the test again and aida locked up. Windows was still responsive, so I just shut down the system at that point. Is aida that much worse it can detect an error almost instantly, that 20m+ in memtest wont show up?

What's a good general strategy for testing ram OC? I've heard about bricking Windows so I'm wanting to proceed with caution here. Also, anyone know what kinda ram speeds Ryzen can get up to? Or I could just hold off until I move it into an Intel system...


Edit:

I tried relaxing it to 16-16-16-38 still at 3200, but that failed aida after around 20 minutes.
Going back a step, I went to 14-14-14-34 at 2933, which is the best setting not exceeding the ram rating, and that passed 70 minutes of aida.
I went into the bios to see what I could tinker with. On intel systems, I might look at VCCSA and VCCIO but I didn't see those. Absent anything else to try, I raised the ram voltage to 1.4 and upped VTT to 0.7. This looks promising, in that I've run 3200 14-14-14-34 for a few minutes so far without error in aida. Nope, failed after 7 minutes. But still a big improvement over failing instantly. More voltage? Or is there something else I should tinker with?

Maybe I should move it into my i3 Intel system which isn't active at the moment...

Edit 2:

I've moved the ram into my i3-6100 on MSI Z170A Gaming Pro. I replaced the non-K OC bios with latest official release. Just setting XMP and reverting the bus back to 100 (otherwise it fsb OCs to try to reach target ram clock) it seems to work fine at resulting 2933. Going to 3200 without changing settings seemed ok, but I note the bios behaviour on auto is to also relax primary timings for you, so I was at CL15 now. This ran for 9 minutes of testing before I moved up a notch to 3466. I noted it relaxed to CL16 here. Now aida failed after 30 seconds. Back in bios, I was about to undo it when I noticed VCCSA and VCCIO were both set what I thought was really high, around 1.25v and 1.275v, I forget which was which. Could that be excessive? I dropped both to 1.20v and aida failed after 11 minutes. Possibly a big improvement. Bit late and bed time for me now, so no more tinkering for now.

 

[batboy]

I tend to be conservative with voltages, but I routinely use 1.20v on VCCIO and 1.25v on SA. Some folks go much higher, but these settings can sometimes cause problems when overvolted. I downloaded Aida, but never bothered to learn it. I don't like the demos with time limits anyway.

 

[mackerel]

What do you use to test stability then?

The time limited aida demo was annoying, since you didn't even get to try full functionality during that time, but regardless I did get a licence a while back and still run that.


Edit:
3600 C17 SA 1.25 IO 1.20 failed after 37 minutes
3700 can't get to boot at all
3600 C16 now messing around with voltages to try and find stability

 

[Tír na nÓg]

Don't know about z170 boards, but I noticed on my X99's (Extreme 4 and now Fatlity Pro), SA tweaking is quite tricky: I run the ram as shown in sig. SA offset+0.175v. If I set it above +0.200v, one ore two sticks are not recognized at boot... When pushing to 3300CL13, needs +0.200'ish. If I set the SA offset @+0.250 or more, same issue as above.

Regarding stability test, well, I uise Aida for both CPU and DDR testing for 6 hours (overnight, short nights ).

What I noticed is that RAM instability usually shows up very quickly (lock up, freeze, blue screen, and first sign is "the application blah blah did not close correctly" (or something like this) at system shutdown.

 

[mackerel]

So far I've had some no POSTs, if it gets past that then all but one time I get to Windows desktop (once I got no boot device). Stability then is interesting... running only Aida64 stress test, I can have errors in one of three ball parks: errors quickly (less than a minute), errors short term (up to 10 minutes), errors long term (over 30 minutes). Obviously tinkering when errors might take >30 mins to show up isn't fun.

Anyway, I've given up tinkering at 3600C16 for now. No combo of voltages tried so far will stabilise that beyond a minute or so of stress. I've tried going up and down a bit on both SA and IO, and also tried giving a touch more Vmem too, which if anything might have made things slightly worse.

I've verified rated 3000C14 on the system, as I hadn't done that up to now, passing 1h40m+. Debating what to do next. I can try stabilising 3600C17 although part of me says, if I'm going to do that I should have paid a little more for the 3600C16 kit instead of the cheaper 3000C14 I have here.

How much is the stability dependant on the mobo? I'm using an MSI one which I've never been happy with when it comes to higher speed ram. It has struggled with Corsair 3000 and G.skill 3333 in the past.

Or maybe I can try dropping latency at stock speed? So many settings, so little time

 

[Tír na nÓg]

Yep, tweaking a notch every 45 minutes sucks big time... I wonder where is Woomack?

I was there yesterday, tryimng to bring the rig back to 4.8 (from 4.75, lol). Real Bench was giving error after 40 to 50 minutes...

I gave up!

 

[mackerel]

Decided to try 2933C13 after all (13-13-13-auto). So far passed 10 minutes.

I should add, when I said earlier I verified 3000C14, I was in fact testing 2933C14 due to how the multipliers work out. In latency terms, 3000C14 to 2933C13 is only 5% faster, so hopefully this is an easy gain.

 

[batboy]

I use the memory subscore of geekbench 4 to measure RAM performance (it has two scores, single core and multi core). Admittedly, this does not test for errors. I don't worry too much as long as it finishes the benchmark. If you need to handle mission critical data or if it's a matter of life or death, then you probably shouldn't be overclocking.

 

[mackerel]

I'm ok with the notion of "bench stable" but I've tested and run for 24/7 OCs too. I tended to do low voltage OCs to get easy untapped performance and thus my eventual clocks tend to be far lower than everyone else. I don't do it for ram as that generally speaking didn't have enough gain for the risk and effort. Even if this ram is primarily for benching, I'm also curious how stability works and changes.

2933C13 has passed 64 minutes now. Do I try for 12?

 

[Tír na nÓg]

^Of course!

 

[mackerel]

2933 12-12-12 no boot
2933 13-12-12 under testing

Edit:

2933 13-12-12 passed 65 minutes

Maybe 13-11-11 next?

 

[Tír na nÓg]

TRy working on the sub timings (I found tRFC has quite a big impact).

 

[EarthDog]

Aida64 is 'multi' threaded and scales with the number of threads on the cpu while memtest doesnt change with the number of cores/threads.

I would stay away from using geekbench results though... its not really made for testing that and its scoring can be i consistent.

 

[mackerel]

I'll try tRFC later.
The version of memtest I'm using seems to default to using all cores. It was certainly still an option on older versions. Not sure that explains the difference.

Anyway, I got distracted so the 13-12-12 test actually passed 2h48m before I stopped it.



No idea if this is any good, I'll have to find something to compare against later. Gotta go do other things for now.

 

[Woomack]

Memtest is not checking cache and IMC stability, it's just checking cells one by one, in new/multithreaded version it's checking it faster ... that's the short description. AIDA64 in memory/cache test is loading 95-97% of RAM, regardless if you have 8 or 64GB, it will be always about that. It tests almost all RAM and you still can make screenshot or open additional test.

Easy way to OC B die is to start at 2133, write down timings, set 3200 and use the same timings. It works on most platforms. Good sticks will run even higher.
On Ryzen there are not many timings and not everything will work so if you can set 3200+ depends on manufacturer's timing profile which in big part is not visible to users and can't be changed. This is why some kits are working and some not even though all can be on the same IC. In general it's easier to set higher clock on 3466-3800 kits on Ryzen because of more relaxed sub timings ( which you can't see ) but there are sometimes issues with 4000+ kits. For me the best working on Ryzen was Ballistix Elite 3466 kit.

 

[mackerel]

Bit late now, kinda wondering if I should have got the 3600C16 kit instead of the 3000C14. I'm winding down my other activities so should free up the Coffee Lake system for testing from tomorrow. Maybe.

 

[Woomack]

I just sold 3000 C14 and 3600 C16 kits... both were on the same IC and were overclocking about the same on Intel. 3600 C16 kit could run at a bit tighter timings at 3733+ but I mean for quick benchmarks at CL12, not 24/7 stability. Maximum clock on both was about the same ~4266+.

On Ryzen I would run that at 3200 14-14-14 or 16-16-16 1.35-1.40V depends on motherboard ( performance will be about the same ). In tests tighter timings are not really helping. I guess that AMD technical department and marketing are not working with each other as AMD is recommending memory kits which are or worse or are not working at all on their platform.

 

[mackerel]

I saw your previous write ups on both kits. I still wonder, how much of it is down to the mobo (+bios) vs the modules. Am I not getting the higher speeds you got because I don't know what I'm doing? Very possible! Or is there a mobo limit? At a first level, it is adequate to play with primaries and voltages? Or do I need to start looking more at other timings already? I was unable to find stability although it was more than bench stable at 3600C17 for example, but not at C16. With my past testing, I recall finding a sweet spot for both VCCSA and VCCIO, where going above or below the "best" level lead to reduce stability and/or more detected errors. But in reading around, I'm getting the impression SA helps influence most higher clock stability, and IO is less important. But I'm still lacking the experience to get a feel for what sort of ball park levels are sane here.

When buying this, part of my thinking is that I'm doing things differently than I normally do. I normally just want enough ram bandwidth, don't care about latency. This kit might allow me an easier starting point to explore lower latency, regardless how much it might help. The other part of me says the 3600C16 kit might be easier to use outside of benching, if XMP works. I'm really not convinced on general mobo auto functionality at such speeds. To help persuade me, I thought the 3000C14 kit expensive enough, as there seems to be a premium to get anything B-die. The 3600 kit was another step above that in pricing...

 

[Woomack]

3600 16-16-16 works at 1.35-1.40V on most b-die. Better kits will make 15-15-15. There are couple of reasons why it may not work stable. In general on KL/CL it should run at 3600 at stock IO/SA voltages but you can set 1.05-1.15V to be sure.

There are no special settings up to ~3600. Can leave almost everything at auto. Check if XMP +manual timings is working or full manual. When you leave XMP enabled but set everything else manually then whatever is left at auto should be adjusted a bit different ( like in XMP ).
CR2 is sometimes helping.

If you test it on coffee lake then check the latest BIOS. Some users have problems on early releases.

 

[mackerel]

In general on KL/CL it should run at 3600 at stock IO/SA voltages but you can set 1.05-1.15V to be sure.

With XMP on, regardless of the clock, the MSI board autos IO to around 1.25 and SA 1.275. Too high? I didn't go that low in testing.

No new bios (since launch) for my Z370 board yet.