F4-3600C16Q-64GTZNC not stable at 3200MHz and beyond on QVL motherboard

[DanielPharos]

Over the past couple of months, I've been having difficulty getting my G.skill memory to run properly.

My set-up:
AMD Ryzen 3800X
ASRock X570 Steel Legend
G.skill F4-3600C16Q-64GTZNC
Zotac GTX 1080 Ti

I bought this set-up so I can run the memory at 3600MHz with the IMC in 1:1 mode, for the lowest latency without having to risk the silicon lottery for 3733MHz. I want to be able to run several VM's while playing a demanding video game, hence the large amount of memory.
Since my motherboard is on the memory's QVL list, I'd expect to be able to run the RAM at its advertised XMP speed. However, every time I try to push beyond 3133MHz, Memtest86 (I'm using version 8.3) starts finding errors, and I indeed experience instability.

I've tried many things trying to stabilize my build, learning a lot about DDR4 and how Ryzen's work along the way. Here's the latest procedure I've gone through:
- Clear CMOS. (RAM defaults to DRR4-2133.)
- Set FLCK to 1800MHz. Reboot, test, works fine.
- Set CR to 1T (with GearDown disabled). Reboot, test, works fine.
- Turn on XMP, and turn the memory up to 2800MHz. Reboot, test, works fine.
- Crank it up to 3000MHz. Reboot, test, works fine.
- Crank it up to 3166MHz. Reboot, test, works fine.
- Crank it up to 3200MHz. Reboot, test, errors.
- 3200MHz, but with tCL set to 24. Reboot, test, errors.
- Cranking it up to 3600MHz results in boot failures and heavy memory corruption.

If I just try all-Auto settings, it goes through a bunch of training failures, ending up with the FCLK at around 1600MHz. However, it's not actually stable: stressing the set-up results in memory corruption (crashing games, crashing Windows, and Memtest86 can sometimes find it as well).

I've already gotten a new RAM kit once, and a new CPU once. I started out on BIOS version 2.20, but upgraded to 2.30 a couple of days after its release. None improved the situation.

So I need some help. What should I do to get my G.skill memory working as intended?

[DanielPharos]

Any news on this? I still haven't been able to get it stable.

[mta]

Just from my own experience: I have to keep Geardown Mode enabled at those upper frequencies - often times I can't boot without it on. Maybe try again with that setting enabled instead.

[DanielPharos]

I've tried that already, without success. In one of my attempts, I forced Geardown Mode on throughout the entire tuning-reboot-increase cycle, and it didn't result in stability at and beyond 3200MHz. I also did some attempts with Geardown Mode set to Auto, and disabled; I never got any stability at and beyond 3200MHz.

[HisEvilness]

Hello,

While I was working on my X370 board with QVL and none QVL memory from Corsair and G.Skill I had similar problems what you want to do is the following.

1. SoC Voltage.
2. ProcODT.
3. Gear Down Mode.
4. CLDO VDDP Voltage.

Those need to be adjusted but they depend on the kit etc so I would not know the exact settings. I wrote down my workflow here: https://www.hisevilness.com/articles...ty-fabric.html Some of the voltages and save voltages are linked in top of the page in a separate article.

Hope that helps.

[DanielPharos]

Much appreciated! Unfortunately, I'm not sure how useful your information is in my case. I'm on Ryzen 3000, which has different characteristics and frequencies. Additionally, your step-by-step plan doesn't force a 1:1 coupled mode, which is something that is absolutely required on my setup; by letting it DDR-train on start-up with higher frequencies, it will quickly fall down to 2:1 (= 2T command rate), which is exactly what I don't want. Also, it automatically starts lowering the Infinity Fabric frequency, which is also something I explicitly don't want. Hence me forcing the Command Rate and FCLK in the first few steps of my written-down attempt.

I already tried adjusting the SoC voltage (that's what I started doing initially), but my MB+CPU combination already defaults it to the maximum safe 1.2V. When I played around with the ProcODT, it didn't seem to affect stability. Gear Down mode doesn't affect stability for me either. And I played around with the VDDP voltage, but I read conflicting information about whether it should be higher or lower than the default, so I tried both, and neither helped increase stability.

It's quite concerning you also had trouble getting QVL memory working properly with your set-up. Is Gskill perhaps not testing stability correctly in their QVL tests?

[HisEvilness]

These settings are different per PC setup and then kit. I've done 2 so far(none QVL) as they are listed I bought another Kit that is QVL(F4-3200C14D-16GTRS). It is not the memory fault it is the infinity fabric that is unstable in essence. Important to note that Gear Down Mode only works with CR1 you should see it as CR 1.5 and will lock out CR2. Gear Down Mode works best running higher frequencies where the infinity fabric can not keep up. Running out of Sync is bad yes you should definitely not do that. As far as the 3000 series it should be easier than the 2000 or 1000 series with a more fleshed out IMC. As you stated you can pass memtest at lower frequencies so you can be certain it is not a faulty memory stick. CLDO VDDP voltage is to patch memory holes most of the time is should be lowered but at times it runs better with higher mv(would indicate a lack of memory holes)Since your are on QVL memory you can either RMA it and get another kit, i am interested to know if you happen to know the ICs on the sticks? Also running lower frequencies with tighter timings tends to be better for throughput of the memory as well as lowering the latency. Example here: https://imgur.com/a/pgiLe#hf4B4N7

Again I deliberately tested none QVL kits this is not G.Skill fault. The QVL i have runs like a charm

[DanielPharos]

These settings are different per PC setup and then kit. I've done 2 so far(none QVL) as they are listed I bought another Kit that is QVL(F4-3200C14D-16GTRS). It is not the memory fault it is the infinity fabric that is unstable in essence.

Yes, I figured out that the instability is somewhere in the chipset+CPU, not the RAM, because using fewer memory modules makes the set-up much more stable.

Important to note that Gear Down Mode only works with CR1 you should see it as CR 1.5 and will lock out CR2. Gear Down Mode works best running higher frequencies where the infinity fabric can not keep up. Running out of Sync is bad yes you should definitely not do that. As far as the 3000 series it should be easier than the 2000 or 1000 series with a more fleshed out IMC. As you stated you can pass memtest at lower frequencies so you can be certain it is not a faulty memory stick.

That, and I already had one kit replaced, as I stated in my opening post.

CLDO VDDP voltage is to patch memory holes most of the time is should be lowered but at times it runs better with higher mv(would indicate a lack of memory holes)

I've already tried both lowering and raising it, but it's difficult to find what the nominal value should be in the first place, and what the safe ranges are. Do you happen to know these?

Since your are on QVL memory you can either RMA it and get another kit,

(Already did that once; I'm pretty sure that's not going to help?)

i am interested to know if you happen to know the ICs on the sticks?

Where do I find this information?

Also running lower frequencies with tighter timings tends to be better for throughput of the memory as well as lowering the latency. Example here: https://imgur.com/a/pgiLe#hf4B4N7

But when running at lower frequencies in 1:1 mode, the entire Infinity Fabric runs slower. Doesn't that decrease overall system performance too?

Even so, G.skill says this memory should be able to do 3600MT on this MB, and I specifically bought this entire set-up for this, so why doesn't it work (for me)?

Again I deliberately tested none QVL kits this is not G.Skill fault. The QVL i have runs like a charm

Why are you jumping to conclusions that this isn't G.skill's fault? *They* claim my set-up should work, and it doesn't. It's very likely I'm doing something wrong, and that's why I'm here, but to presuppose G.skill is not at fault at all seems like an unwarrented assumption to me at this point.

[HisEvilness]

CLDO VDDP volt for me is in the 900 range and I've seen it go up as high as 1000 while doing some research on it, information on it is scarce why I have to spend extra money on memory kits to do some field testing so to speak. As far as a good range I am in the 900 to 925 range with 3 kits.

As far as the kits go I might have missed it but it looks like that the CPU IMC can not handle 64GB I can remember my 1600X was not doing so well with 32GB as it put to much strain on the IMC. This, in turn, might make it impossible to stabilize at all. G.Skill tests a number of motherboards and then most likely has a tray of CPUs that their engineers put on motherboards yours might be exceptionally bad IMC wise. I am not sure if money is an issue for you but buy perhaps a cheap 3300X or something that can run 64GB QVL memory and see if you have the same problem.

You can read your stick ICs with a program called Thaiphoon Burner. Uploaded 2 examples so you have an idea what to look for.

Lower frequencies is not always bad IF you can tighten the timings but this requires a lot of work and you just want XMP what is fine. The example was to illustrate that a lower frequency with tighter timings will enhance your memory and overall PC performance. A general rule of thumb for overclocking memory is only to step up the timings by 1 if you gain 200mhz or more.

As far as G.Skill go I do not present them and I was not jumping to any conclusion merely pointing out that it should work as you would expect but there are outlining cases and it looks like you are one of them where due to circumstances QVL memory is not working. I had the same with my 1600/1600X I spend weeks looking for a good kit that would not randomly shut down my PC and was very frustrating so I can understand your disappointment.

[DanielPharos]

CLDO VDDP volt for me is in the 900 range and I've seen it go up as high as 1000 while doing some research on it, information on it is scarce why I have to spend extra money on memory kits to do some field testing so to speak. As far as a good range I am in the 900 to 925 range with 3 kits.

You're not using a Ryzen 3000, are you? I've seen differences in default voltages between Ryzen 1000-2000 and 3000. My default is 1.1V (according to the AMD Ryzen Master utility).

As far as the kits go I might have missed it but it looks like that the CPU IMC can not handle 64GB

This is what I suspect as well, meaning that the QVL-rating for this memory kit is misleading.

I can remember my 1600X was not doing so well with 32GB as it put to much strain on the IMC. This, in turn, might make it impossible to stabilize at all. G.Skill tests a number of motherboards and then most likely has a tray of CPUs that their engineers put on motherboards yours might be exceptionally bad IMC wise.

I've already had my CPU replaced (as per my opening post). Also, the 3800X is a binned CPU, so having two exceptionally bad CPU's in a row is very, very unlikely.

I am not sure if money is an issue for you but buy perhaps a cheap 3300X or something that can run 64GB QVL memory and see if you have the same problem.

Yeah, I thought about doing that, but then I figured I'd try asking here before I started throwing money at the problem.

You can read your stick ICs with a program called Thaiphoon Burner. Uploaded 2 examples so you have an idea what to look for.

Thanks for the clear screenshots! Here's the info:

ORGANIZATION
2048M x64 (2 ranks)

REVISION / RAW CARD
0000h / B1 (8 layers)

MANUFACTURER
Hynix

DIE DENSITY / COUNT
8 Gb D-die (Davinci / 17 nm) / 1 die

COMPOSITION
1024Mb x8 (64Mb x8 x 16 banks)

Lower frequencies is not always bad IF you can tighten the timings but this requires a lot of work and you just want XMP what is fine. The example was to illustrate that a lower frequency with tighter timings will enhance your memory and overall PC performance.

Aren't the PCI-lanes and other IO also going through the Infinity Fabric? I have no desire worsening performance for those just to fix some instability that shouldn't be there in the first place.

A general rule of thumb for overclocking memory is only to step up the timings by 1 if you gain 200mhz or more.

I'm not looking into overclocking (other than running this memory at its rated XMP speeds, with a MB that's on its QVL, with a CPU-platform it explicitly supports).

As far as G.Skill go I do not present them and I was not jumping to any conclusion merely pointing out that it should work as you would expect but there are outlining cases and it looks like you are one of them where due to circumstances QVL memory is not working.

If that's the case, it shouldn't be on the QVL list.

I had the same with my 1600/1600X I spend weeks looking for a good kit that would not randomly shut down my PC and was very frustrating so I can understand your disappointment.

If that's what G.skill thinks is acceptable for QVL...

[HisEvilness]

Correct I am not on Ryzen 3000 I am using Ryzen 2000 and 1000 but the only difference would be the CPU voltage not the CLDO and other voltages.

To sum up several quote sections, binned chips should be good to go, replacement QVL kits were used as well. And looking at the specs I never heard of D-die but it is rather new and it is an improved CJR(C-die) and that is defiantly Ryzen friendly. Binned Ryzen chips also make it highly unlikely your IMC is bad at all. The only thing I can deduct then is your motherboard might be the problem. You can test it but this will be tedious it means you have to memtest 1 stick per slot and make sure the stick you use is stable. And then memtest 2 sticks per channel. X570 is dual channel, so the first 2 slots are channel A and the last 2 slots are channel B. The kit you have is dual rank, in short ICs on both side of the PCB but with recent updates and work on the Zen architecture this should not be the problem either.

Tighter timings are fine and you have them with XMP so it would not require any work at all, https://www.gskill.com/product/165/2...V64GB-(4x16GB) There is a TridentZ RGB and TridentZ RGB Neo variant of that.

Overclocking wise it is not for everyone I understand I was merely pointing out the gains in tighter timings perhaps I worded it wrong my apologies for the confusion. And the memory kits on 1600/1600X were from various vendors and it was driving me up the wall tbh not G.Skills fault. What you get for pre-ordering 1st gen product they do tend to have their teething problems.

[DanielPharos]

Correct I am not on Ryzen 3000 I am using Ryzen 2000 and 1000 but the only difference would be the CPU voltage not the CLDO and other voltages.

I vaguely remember seeing different numbers, so I'm not too keen on changing them too much without knowing the ranges.

To sum up several quote sections, binned chips should be good to go, replacement QVL kits were used as well. And looking at the specs I never heard of D-die but it is rather new and it is an improved CJR(C-die) and that is defiantly Ryzen friendly. Binned Ryzen chips also make it highly unlikely your IMC is bad at all. The only thing I can deduct then is your motherboard might be the problem.

True, that's one of the explanations left. However, broken how? I'd expect quite different behavior from a defective motherboard. For example, the errors reported by Memtest86 move around, but (almost) always appear in pairs, and appear to move between the different modules and banks. They survive soft reboots, but not full power-cycles. Swapping modules around doesn't change (in)stability. Each individual memory module (in the appropriate slot) passes Memtest86 and other tests just fine.

With the IMC being part of the CPU, there's not a whole lot of MB in between the CPU and RAM, and this is clearly not a case of a broken RAM bank/slot.

You can test it but this will be tedious it means you have to memtest 1 stick per slot and make sure the stick you use is stable.

This I already did, with both my old and replacement RAM.

And then memtest 2 sticks per channel.

This I didn't do in all possible combinations, as I didn't see the value in doing so. If the modules individually are fine, are sold as a single kit, and the manufacturer tested them, what would a failure in such a test point to?

X570 is dual channel, so the first 2 slots are channel A and the last 2 slots are channel B. The kit you have is dual rank, in short ICs on both side of the PCB but with recent updates and work on the Zen architecture this should not be the problem either.

Well, there appears to be restrictions. If you look at various QVL's, you'll see that double ranked memory often only is supported with fewer modules, as the IMC has difficulty driving that many channels at higher frequencies. However, no such limitation is indicated by G.skill.

Tighter timings are fine and you have them with XMP so it would not require any work at all, https://www.gskill.com/product/165/2...V64GB-(4x16GB) There is a TridentZ RGB and TridentZ RGB Neo variant of that.

Yeah, I was looking at that new RAM, but I'm not sure what the difference with my current RAM is?

I did however spot some new Crucial RAM that spec-wise matches my current RAM, but only has 2 modules. If I'm going to switch RAM, I'd like to try that first.

Overclocking wise it is not for everyone I understand I was merely pointing out the gains in tighter timings perhaps I worded it wrong my apologies for the confusion.

I was indeed originally planning in figuring out if I could tighten the timings, but as it's not even stable at advertised speeds, I never got around to trying that.

Also, no need to apologize. I'm just a bit annoyed that, after months of fiddling around, my set-up still isn't running as it should be able to out-of-the-box.

And the memory kits on 1600/1600X were from various vendors and it was driving me up the wall tbh not G.Skills fault. What you get for pre-ordering 1st gen product they do tend to have their teething problems.

I understand, but I'd have hoped that by now G.skill would have figured out and resolved these kinds of issues (although I understand Zen 2 is different from Zen 1 in many respects).

[HisEvilness]

Since you tested out different RAM kits and different CPUs all of high quality, then unless you missed something it can be the motherboard having a faulty/damaged DIMM slot or damages/faulty traces. Memtest errors can be because the signal moving through the mobo via the traces and then into the DIMM slots to the memory with possible damage to the motherboard would introduce corruption. If you know some with a motherboard that can handle this kit perhaps test it out there? If your CPU + Memory runs fine on any other motherboard it would confirm your current motherboard is the problem.

As far as dual-sided memory I never use them since they make for poor overclocker I prefer single-sided. If I were to recommend a kit single-sided B-die with a capacity of your liking. The best kits are 3200C14 @ 1.35, 3600C15/16 @ 1.35v. Running 4 sticks in dual-channel is not a huge problem. I've got 4 sticks running 4000C17 on my Intel rig and it runs like a charm. With a comparable dual-sided Micron E-die kit doing far worse and struggling.

The linked kit is most likely Samsung B-die due to the timings + frequency. Not sure what the Crucial kit is but there is a option to browse kits in Thaiphoon burner to get an idea.

[DanielPharos]

Since you tested out different RAM kits and different CPUs all of high quality, then unless you missed something it can be the motherboard having a faulty/damaged DIMM slot or damages/faulty traces. Memtest errors can be because the signal moving through the mobo via the traces and then into the DIMM slots to the memory with possible damage to the motherboard would introduce corruption. If you know some with a motherboard that can handle this kit perhaps test it out there? If your CPU + Memory runs fine on any other motherboard it would confirm your current motherboard is the problem.

I unfortunately do not have a spare MB around, or I would have tested this already.

But before I start ripping this entire build apart again, I'd like to know if this set-up even can work.

As far as dual-sided memory I never use them since they make for poor overclocker I prefer single-sided. If I were to recommend a kit single-sided B-die with a capacity of your liking. The best kits are 3200C14 @ 1.35, 3600C15/16 @ 1.35v. Running 4 sticks in dual-channel is not a huge problem.

Are there any 64GB 3600CL16 options out there, except for G.skill and the Crucial kit I referred to, that are confirmed to work at the advertised speeds and timings with the X570 chipset and a 3800X CPU? If you know of any, please tell me!

I've got 4 sticks running 4000C17 on my Intel rig and it runs like a charm. With a comparable dual-sided Micron E-die kit doing far worse and struggling.

I am aware of some RAM not being able to do this, but G.skill advertises this RAM as capable, so it should be. And Intel's memory controller is quite different, so that it is capable of some cool feats like that isn't surprising to me, but I'm not switching my set-up to Intel just to get this RAM operating properly.

The linked kit is most likely Samsung B-die due to the timings + frequency. Not sure what the Crucial kit is but there is a option to browse kits in Thaiphoon burner to get an idea.

(I'm not going to mention the exact Crucial kit; I'm not sure that's appropriate on the G.skill forum.)
Wait, one can browse memory kit specs in the Thaiphoon burner application? I'll check that out; that sounds very useful!

[HisEvilness]

It should work why i point out that the motherboard might be at play but correct you would need spare parts or have a friend who has this laying around.

QVL for your motherboard is listed here: https://www.asrock.com/mb/AMD/X570%2...x.asp#MemoryMS It covers most brands just make sure you use C16 with 64GB capacity. As far as i can tell looking at the page for Mattise there all Samsung B-die.

My 4000C17 is just to illustrate that using 4 sticks is not a problem the same kit is also on the QVL list for Ryzen. Base speed is 3466 not the best bin but it is right up there.