Z170 RAM expertise needed (trying to get stability)

[superstition]

(Please don't invoke the debate about the gains to be found from faster DDR4. I want to focus specifically on getting stability in a speed range of 3200–3600 and need specific, clear, data-driven info. Thanks!)

I am setting up a computer system for a friend and bought a 16 GB (2 X 8) EVGA 3200 1.35 pack (Samsung chips, dual rank I believe) and installed it into an AsRock Z170 Gaming ITX/AC board. It has the latest BIOS.

stock: 16-18-18-38-2T-540, 3200, 1.35
mine: 15-16-16-36-1T-460, 3466, 1.385 (power supply overvolts 3.3V line to ~3.360), 1:26

6700K:

stock: 1.23V
mine: 1.15V, level 2 LLC (voltage seems rock solid under Prime so far, definitely the least variation at 1.15 in BIOS under load — basically none)

I thought I had stability because I did Prime (including 90% RAM) for two hours and an hour of MemTest (8 copies in Windows). However, I ran MemTest again and now every time I run it I get about 200% coverage of the 2048 MB parts and then one of the tests with the smallest memory increment gets an error. The first failure was with a 768 MB partition. The next was in a 256 MB partition. The third was in a 768 MB partition. The 2048 MB and 1024 MB partitions never error.

My hypothesis is that the RAM is heating up in those smaller partitions more. Part of the trouble is that the ITX case I have requires that my H90 cooler blow its heat right toward the RAM. So it does run rather hot.

It's possible that turning off fast boot and allowing the RAM to train at boot may be causing, possibly just in part, the instability. But I can't say for certain because I didn't run MemTest long enough with small enough RAM increments.

As usual, there are wildly conflicting reports about voltage and DDR4. Some sites use really high voltages, making their results useless for me to try to work with because they have lopsided settings like 14-17-17 at very high speeds that probably require lots of voltage. (Sadly, these tend to be the only sites that work with the timings I'm particularly curious about, like tWR, tRFC, etc.) Some posters argue that anything about 1.35 is super-dangerous and shouldn't be used. So I backed my RAM in BIOS to 1.33 to get it to 1.35 according to AIDA64 and HWINFO64.

I just bumped the tRFC to 480 to try to get stability but I would like to know more precisely how to determine what a safe tRAS and tRFC number is for 15-16-16-1T at 3466.

I've seen silly claims in various "guide" forum topics that it has to be 41 +-2 (tRTP + tCL + TRCD)which would make a minimum of 39 but that doesn't match actual shipping RAM sets including mine. Its specs are 16-18-18-38 — higher latency settings but lower tRAS! I find it hard to believe that memory manufacturers are shipping most, if not all, of their DDR4 RAM sets at a tRAS below the minimum required.

Input on the timing settings to tweak to get better stability without throwing performance out the window are very appreciated.

Some questions:

• Is bumping tRFC to 480 from 460 enough of a change or do I need to do more to try to get stability?

• Will reducing voltage to 1.35 possibly make me stable, if RAM overheating is the cause of the problem?

• Should I turn quick boot on, disabling training? People say to manually lock in settings but there are a ton of them so that's not exactly a simple proposition. Will leaving training on lead to instability because of different settings each time the system boots? How can I lock the settings if I don't know what they should be — just leave most of the arcane ones on AUTO?

• Should I increase tREFI? One article I read said bit flips can happen with the very board I have above 19,000 so I think keeping it below that is a good idea. It's at 11,000 something currently which is the default. People have said tREFI increases can increase performance but never explain by how much or what the drawback might be.

• Are the really lopsided settings (e.g. 14-17-17) I see in the high voltage articles worthwhile even at 1.35? Are non-lopsided settings better (e.g. 16-16-16)?

• One article used 2T for its high-speed RAM testing. Is it worthwhile to sacrifice the latency under 2T to get better settings in other areas, leading to an overall gain? Everyone else goes to 1T but I see that 2T is popular with RAM makers.

• What is the real proven basis for calculating tRAS? Please give this method in simple numbers not arcane nanoseconds that I can't input into BIOS.

• What is the real proven basis for calculating tRFC? (Simple numbers.)

• Is there any data about Skylake being able to handle anything above 1.35 and below 1.4V?

• Is there a good second test for RAM stability after running 8 instances of MemTest? The whole "just use your PC and see" method won't work because I don't want to give him an unstable system. Also, some have said even using stock settings with high-speed RAM sometimes leads to instability so the answer "Just set it at stock" doesn't necessarily work. I am probably already going to have to wipe and reinstall Windows or roll back to an earlier point because when I booted today some icons were missing from desktop shortcuts which suggests corruption.

• Are my tests flawed because of the heat issue — do I need to blast the video card to heat up the case?

Thanks!

 

[superstition]

Well, I just ran it for four hours or so and got an error in a 768 MB area again. The 256 MB area was fine at 41553% coverage. The 2048 MB partitions were at 710% coverage.

The only thing that I had changed, other than lowering the voltage to 1.33 (1.35 in Windows) was raising the tRFC from 460 to 480 and I increased the tREFI to 16,000.

Not sure what to try now. I don't have infinite time to test. It seems like I'm right at the edge of stability but I don't know if it's the CPU voltage or a specific timing that's to blame.

 

[Dasa2]

memtest86 isnt much use these days for overclocking it takes to long to find errors its only useful for diagnosing seriously faulty sticks of ram

when overclocking my memory i found some useful info in this thread
http://www.overclock.net/t/1569364/...broadwell-e-24-7-ddr4-memory-stability-thread
the linux stress test finds ram errors from overclocking really fast
the hci memtest is excellent for finding memory controller and cache related stability problems

people are running ~1.8vdimm short term for benchmarking some have been running ~1.45v for a while 24\7

in the end i have settled on 3866c16 2t tRFC 310 vdimm 1.425v io 1.15v sa 1.2v with single sided 2x8g samsung b-die in maximus viii gene

<dropped tRFC 10 points since that screenshot it seemed to become unstable in hotter weather

 

[superstition]

the linux stress test finds ram errors from overclocking really fast
the hci memtest is excellent for finding memory controller and cache related stability problems

I've been testing with the latter. But, I tried using System Restore to get rid of any corruption and it corrupted Windows. I put the RAM back at stock before doing it. I'm furious and am very close to giving up and letting the system run it at 2133. I've had nothing but problems with every system I've worked with since the E2140. Nothing ever seems to be totally stable.

Bah. I'm done with this nonsense. If it has to be extremely slow to be stable so be it.

 

[Dasa2]

tRFC to low is a quick way to corrupt your os which is why you should stress test with linux first then make sure everything else is stable with hci
leave sub timings at default for a start find a stable speed around 3200 and fine tune later

 

[superstition]

You're right. It makes better sense to test in Linux with something light like Tails so there's no worry about corruption. Is there a good utility similar to HWINFO64 for monitoring?

I set up a Windows 8.1 installation just for testing with my AMD setup but I didn't feel like going through that, including installing a sata ssd in the ITX case. It's just annoying, frankly, to be stable for hours and then get an error.

tRFC set too low should cause an error sooner. As I understand it it is one of the settings that pretty much either works or doesn't. The "training" BIOS stuff is also a big question mark. I can find all the stable settings I try to but if the automatic timings don't "train" correctly with every boot then I'm wasting my time. One site said to lock them in manually but that means figuring out every setting.

 

[BonzaiDuck]

There was a time when I invested a lot of effort and a learning curve on tweaking RAM.

The last OC'd RAM I'd used was fairly easy to OC: a DDR3 pair of G.SKILL kits (all four slots!) of DDR3-1600 RipJaws OC'd to 1866 and loosened timings. I don't like doing the test in Windows, inviting someone to tell me I'm naïve or something. I use the self-booting CD of HCIMemtest64.

To test a brand new kit so I stay within the 30-day RMA window, I may run 300% coverage at all the stock settings. If, with the same kit, I want the same settings but with CR=1, I might go 500%. But if I'm going to tweak the timings, I might wait days to get 1,000%.

Sure I can test in Windows with multiple instances as someone above mentioned. I can also run certain Prime95 or LinX tests which will reveal anything or suggest bumping up the IMC voltage. Usually, a stop code will suggest VCCIO as a possibility as well as VDIMM or VCORE.

But why spend all this time and effort to achieve those clocks and timings, when you can probably just find the RAM offered with the desired XMP spec, buy it, install it, test it, and fahget-about-it?

How much time would you be willing to save by spending a little extra?

 

[superstition]

We'll see if the XMP profile is stable now that I had to reset the CMOS because everything was going haywire, including the stock XMP, after I tried to switch it on to replace the 2133 BIOS default.

I have a feeling I should have reset the CMOS after updating the BIOS way back when.

The frustrating part is getting to 733% coverage and then getting an error — without knowing what to tweak to fix it.

 

[Dasa2]

The frustrating part is getting to 733% coverage and then getting an error — without knowing what to tweak to fix it.

running the linux test first helps narrow it down
if its linux failing you probably need extra vdimm or slow the ram
if hci fail's after linux its probably sa\io

my mb had problems with training when running 1t over 3600 if i swapped it from 1t to 2t for example i would need to disconnect the psu from the power for a few mins before it would post again then it would be fine
setting everything manually didn't help it was training hidden timings

 

[superstition]

RAM voltage is unlikely to be the issue because I had a longer run time without an error in HCL after reducing the voltage to 1.35 (from 1.38). What is "sa/io"?

Perhaps it's my CPU voltage being too low?

 

[BonzaiDuck]

We'll see if the XMP profile is stable now that I had to reset the CMOS because everything was going haywire, including the stock XMP, after I tried to switch it on to replace the 2133 BIOS default.

I have a feeling I should have reset the CMOS after updating the BIOS way back when.

The frustrating part is getting to 733% coverage and then getting an error — without knowing what to tweak to fix it.

Just reviewing your initial post again, I see you didn't want to hear a "pros-and-cons" about "gains to be made over 'faster' DDR4." My point wasn't about the gains, though; it was about the benefits and costs of OC'ing a large RAM kit. I'm just guessing that you're using multiple instances in Windows to test it, because 733% takes a lot of patience otherwise.

Also, why are you tightening the timings while raising the speed? If it were me, I'd try and keep the same timings at the higher (3,466?) speed, and then just tweak the voltages. Or I'd simply loosen the timings and attempt to keep the same voltages before adjusting further. I'm not familiar -- nor did I know of -- EVGA RAM kits. You probably bought them for the "black parts" since you mention Samsung.

 

[Dasa2]

What is "sa/io"?

System Agent Voltage & CPU VCCIO Voltage
there for the memory controller on the cpu


BonzaiDuck hci windows completes much quicker with a overclocked ddr4 system
8 instances at once 1000% each only took a few hours while i went out that or i left it run over night cant remember which

 

[BonzaiDuck]

System Agent Voltage & CPU VCCIO Voltage
there for the memory controller on the cpu


BonzaiDuck hci windows completes much quicker with a overclocked ddr4 system
8 instances at once 1000% each only took a few hours while i went out that or i left it run over night cant remember which

Well, I'm humbled by that. I initially tried using the program that way, and I just felt uncomfortable about it. I decided to test with the bootable CD, thinking no chance of OS corruption.

Interesting, how one can miss some common-sense for so long. Next time, I'll think to ask someone else's opinion.

 

[superstition]

System Agent Voltage & CPU VCCIO Voltage
there for the memory controller on the cpu


BonzaiDuck hci windows completes much quicker with a overclocked ddr4 system
8 instances at once 1000% each only took a few hours while i went out that or i left it run over night cant remember which

Four hours or a bit more got me to 710% coverage for the 2048 MB partitions at 3466 speed with a 6700K. It's still a slow process, particularly since it took that long to get a single error.

 

[superstition]

why are you tightening the timings while raising the speed?

I read reviews and looked at what people were getting from 16 GB (8 x 2) sets. While my RAM isn't binned the same and may not even be using the same manufacturer I noticed that the only time one really gets good gains with higher speeds is when one also keeps the timings tight. While 2133 will bottleneck and possibly 2400, 2666 at reasonably low timings is probably enough. But I was hoping to do a bit better than that.

One review using my board and CPU got their RAM to 4000 speed at 1.35V. I thought that if that can be managed then I should be able to get decently tight timings at 3466 or so. And I did except that I have some small adjustment that needs to be made somewhere. Dasa suggested the System Agent Voltage & CPU VCCIO Voltage — neither of which I touched during my RAM work so either or both may be the thing that should be increased a bit.