Threadripper BUILDERS thread

[ub4ty]

Dude, I'm so not sharing as soon as I get my timings stable later this month because of your attitude. First, IMC varies, meaning having a better chip can change this ad is. Second, p95 doesn't tell anything if your OC isn't stable in the first place, not can you always narrow the ram testing to the ram alone being the cause. Third, considering my ram have ran, granted on an Intel setup, 3733@CL14 for about a year, it is down to timings and IMC. Fourth, that is because those two kits are a quad kit. So they didn't grab two of the 2x8 kits. Big surprise.

You're running a 4133 Trident z (F4-4133C19D kit) downgraded to :
3600 CL16 18-18-38 .. . relax and thanks for the disclosure.

From my motherboard manufacturers website :
DDR4 3600 8GB G.Skill F4-3733C18Q-32GTZSW SS 2pcs 4pcs v
https://gskill.com/en/product/f4-3733c18q-32gtzsw
Tested Latency 18-18-18-38

Timings are shared openly by the manufacturers of the motherboards and the manufacturers of RAM who do a depth of testing to ensure QVL. I can find my settings there. I simply wanted to see what was going on here w.r.t to frequent comments about stability whereby motherboard manufacturers found none. The proclamations here could have caused someone to spend more money on ram than they should have and/or caused them to RMA perfectly fine ram. I'm not sure this is realized..

The fact that you just disclosed your'e not running a 3600Mhz rated kit but are running a 4133Mhz kit that's downgraded with custom clocks is all I needed to hear tbqh. Thank you for the information. I'll likely find stability at CL18 (18-18-18-38) because that's what my spec'd ram more closely matches.. A 3773 rated kit not a 4133 rated kit. The again, 3200Mhz @CL14 will perform better for my workflow. I just came back from a haitus. I'm not engaging anymore inflammatory posts. I expressed a tone of disbelief in the information that has been provided which was warranted given that it didn't match a single QVL list for X399 motherboard. That was confirmed once you stated what rating your memory actually is and your settings.

 

[Markfw]

It isn't the end all be all.. relax buddy. Why do you think no Motherboard vender states that you can run 3600Mhz rated ram at CL16-16-16-36 with a 4x8GB configuration? I'll wait for the answer. What you're forgetting here is that Memtest doesn't stress the CPU simultaneously while pushing the RAM. Why do you need to do both? because you have a wattage manager on the CPU, you have two independent dies and you have a cross die infinity fabric that is clocked to the RAM settings in addition to the on-die infinity fabric. So, no motherboard QVL list states that 3600Mhz rated ram is stable at CL16 settings (samsung b-die) even, yet you guys are claiming you have magic sticks that is somehow stable? Simply.. No. It's not stable. Ddog just confirmed my exact observations and this is on the Asrock Taichi X399. So, unless you have a different motherboard, I'd love to hear your extended comments? I'd rather we be critical about this and clear than sit here and get upset at somebody pointing this out and have someone spend $50-100 more on ram that won't run at the rated speeds or we allow for an issue to be raised for a specific mobo manufacturer to look into. As for your'e tone, I'm not getting dragged into another one of those exchanges.

2 people have confirmed 3600Mhz@CL16 (4x8GB) settings on samsung b-dies is unstable on Asrock X399 taichi. This is inline w/ Mobo manufr's QVL list. Period.

Guys. I don't want to start a war, but here is proof that my memory runs@3600 4 STICKS stable@100%load for days.

Edit: Note that I have heard that the RGB sticks had problems where the ones without were fine, I can't remember where I read that.

 

[ajc9988]

Now were talking. This is the detail that needs to be discussed !
Your RAM : 4133 Trident z (F4-4133C19D kits)
Your tmings : 3600 CL16 18-18-38 ..
So, 16-18-18-38 not XMP that usually has it set at 16-16-16-36 for most 3600Mhz kits.
2T not 1T

Calm down your tone btw... You're not even running 3600Mhz sticks. Please indicate where you disclosed this earlier. Furthermore, you're running this at 16-18-18-38 timings not 16-16-16-36. Even further, you're running it at 2T and not 1T. Essentially, you've engaged in detail memory training and found a setting for 4133 sticks that is stable and none of this was disclosed. Meanwhile, to everyone you say I ran my sticks at 3600Mhz. You see why all of your details matter now?


Lots of assumptions and an over-pronounced tone ... I have 4-5 usb sticks of stability testing suites and tools. Memtest is one of them and I know exactly how to probe memory for issues. I just got my setup rebuilt last night (crap AIO cooler that I didn't want to do stress testing on). So, of the tests that I ran, was prime95. Prime95 showed hardware error memory bit flip issues @3600Mhz CL16 (the rating of my ram and the XMP setting). So, I knew right away what was happening. I can confirm this with Memtest but I dont need to. Instead, I thought, first thing : let me post on the thread and see if I am the only one. Now, I could have gone and testing this further myself, done memory training, found a sweet spot that worked for me, and not disclose the settings and then say.. Yeah, hey my stuff is stable at 3600Mhz @CL16. But, what would that have done for everyone else? Were on the same page here AJC, were all trying to stabilize our systems. Lets drop the tone and stick to the technical details.

My observations, QVL list, your comments, and Ddog comments all clearly suggest that, if you want to run @3600Mhz on your memory sticks 4x8GB, you're going to have to do memory training and find a stable sweet spot. 3600Mhz 4x8GB is not stable straight out of the box at XMP profile settings

Now, this profits the whole community knowing this just like knowing that compatible AIO water coolers don't cover the heatspreader accurately and perform poorly on OCs and results in 5-10*C temperature increases over an air cooler that covers the whole die.

My issue was the over reliance on one test. After you explained the rest of your testing, it makes more sense.

My technique is to use Thaiphoon Burner to get the chip numbers, download the data sheet to get voltage limits, resistance, and timings. These are often in ns, so a conversion may be needed. I then look at the XMP timings and convert to ns. Now I have a range of timings to play with. So, between these two timings, you should be able to find something stable, if not tighten latter. I then look for the lower speed kit using the same chips to see what the XMP timings there are, to see if it fully scaled. With this, you have multiple data points to start.

Once I find a stable speed and timings, I validate with multiple tests. After stability is achieved and verified, I then pull out SPi32M to analyze the effects of tightening timings, performing a lesser verification after each timing. After the primary timings are done, I fully validate. Then on to the secondary. Etc. As I said, I cannot fully validate for a couple weeks (work), but that will be forth coming, along with validation and comparison at lower speed with tighter timings. I want it to perform to the best of its ability is all.

 

[Markfw]

This is my ram : https://www.newegg.com/Product/Product.aspx?Item=N82E16820232194

 

[ajc9988]

Guys. I don't want to start a war, but here is proof that my memory runs@3600 4 STICKS stable@100%load for days.

Edit: Note that I have heard that the RGB sticks had problems where the ones without were fine, I can't remember where I read that.

On the Asrock QVL, it shows RGB sticks achieving lower multipliers than is seen in the non-rgb sticks, for one...

 

[Markfw]

On the Asrock QVL, it shows RGB sticks achieving lower multipliers than is seen in the non-rgb sticks, for one...

Maybe this explains the reason mine are stable @rated speed and timings using the (I think this is the name) the XFR profile ? Forgive me if my terminology is askew. Or is that XMP ? whatever the bios has.

 

[ajc9988]

Maybe this explains the reason mine are stable @rated speed and timings using the (I think this is the name) the XFR profile ? Forgive me if my terminology is askew. Or is that XMP ? whatever the bios has.

Xfr = extended frequency range. XMP is extreme memory profile.

 

[Markfw]

Xfr = extended frequency range. XMP is extreme memory profile.

Well, I can't screen copy bios, but its the only non-default choice. I pick it, and it knows its 3600 ram, and I leave it at that. Set the vcore to 1.2, and the speed to 3900, and BAM, thats what I am running. Everything else default.

Bios version is 1.3 or 3.0 or something like that. Whatever the newest was 2 weeks ago.

 

[Chaotic42]

Holy cow, you guys have some beautiful systems. Mine is just thrown together and sitting on the floor. It has a window because it was the best case they have, but it can't be seen and the computer is surrounded by crap.

 

[ddogg]

Maybe this explains the reason mine are stable @rated speed and timings using the (I think this is the name) the XFR profile ? Forgive me if my terminology is askew. Or is that XMP ? whatever the bios has.

I don't have the RGB ones. My symptoms are identical to ub4ty. So I'm not sure why I'm not fully stable at the xmp settings while you are. Either way, not a big deal

 

[ajc9988]

Well, I can't screen copy bios, but its the only non-default choice. I pick it, and it knows its 3600 ram, and I leave it at that. Set the vcore to 1.2, and the speed to 3900, and BAM, thats what I am running. Everything else default.

Bios version is 1.3 or 3.0 or something like that. Whatever the newest was 2 weeks ago.

What's your temps? I noticed at 3.9@1.2V, my bench scores went down. At 1.225, they scaled about in line with 3.8-3.85 scores. But, I also tested pushing the heat just up to 68C on the die. At that temp, I saw current throttle from 188-220 down to the occasional 145-165W. This explains most likely part of the missing score on those that do scale linearly. Until I have the custom loop in, I can't do my full analysis though. But per 50mhz, you should get an additional 40-65 points in CB15, from what I've seen so far. But I have A LOT more testing to verify this to a certainty.

 

[urvile]

That looks awesome. What is with the 1500W PSU though? Are you planning to OC that 1920 to 5GHz?

In his defense I'd probably shoot for 1200 myself. Just between the 2 video cards and the CPU it's about 800w. Figure about 900w with a slight overclock. 1KW is probably to tight for my personal tastes. What's 1.2 to 1.5, building a kick-ass machine might as well buy a kick-ass PSU.

I am currently operating on the principal of MOAR!

optical illusion or is the upper GPU sagging line an inch?

Yeah it's sagging a bit. It actually came out of the box like that. Which is why I initially installed it sideways. The card seems to be slightly bent. Don't ask me why or how.....

 

[ajc9988]

Also, has anyone with the current issue (not temp related) thought to flip the prochot switch on the mb? I am waiting for the custom loop with VRM blocks to do so, but that may get around the limit you see on current.

 

[ddogg]

Also, has anyone with the current issue (not temp related) thought to flip the prochot switch on the mb? I am waiting for the custom loop with VRM blocks to do so, but that may get around the limit you see on current.

Why do you need VRAM blocks for this? I initially played with the idea of watercooling the ram but the budget was already blown out of proportion lol

 

[ub4ty]

My issue was the over reliance on one test. After you explained the rest of your testing, it makes more sense.

My technique is to use Thaiphoon Burner to get the chip numbers, download the data sheet to get voltage limits, resistance, and timings. These are often in ns, so a conversion may be needed. I then look at the XMP timings and convert to ns. Now I have a range of timings to play with. So, between these two timings, you should be able to find something stable, if not tighten latter. I then look for the lower speed kit using the same chips to see what the XMP timings there are, to see if it fully scaled. With this, you have multiple data points to start.

Once I find a stable speed and timings, I validate with multiple tests. After stability is achieved and verified, I then pull out SPi32M to analyze the effects of tightening timings, performing a lesser verification after each timing. After the primary timings are done, I fully validate. Then on to the secondary. Etc. As I said, I cannot fully validate for a couple weeks (work), but that will be forth coming, along with validation and comparison at lower speed with tighter timings. I want it to perform to the best of its ability is all.

Far more testing than i will do and you obviously have a far deeper understanding of this subject. As such, the tone of disbelief wasn't so much directed towards you as it clearly doesn't apply to the depth of research/testing you do. That being said, I've brought out a very constructive convo and details that will help the community. You have provided a wealth of detail, have a wealth of understanding, and do a wealth of testing. Nothing but respect to you as such.

Given your disclosures and others, the wider community now understands things far more and can more calmly try to stabilize their systems. For those who haven't purchased hardware yet, they can do so w/ far more understanding. So, this exchange has profited everyone yet respect is still maintained. Thank you ajc !

Btw - your jab about slapping together 2x8GB kits as a quad config is sustained. I can't knock any truthful statement.. So, with some reflections, I must think on how to best address my instabilities, if its even worth it at this point...

My thinking is : 3200MHz @CL14-14-14-36 is a very solid combination in terms of latency/timings and clock speed. This has been verified to run up to 8sticks of 8GB. 3600Mhz carries higher latency and i'm not sure it scales to 8 sticks. 4 sticks requires some work no matter what kit and some timing tuning. Every extra Mhz bump cost more $$$. So, everyone should be aware of this when making decisions.

 

[ub4ty]

Guys. I don't want to start a war, but here is proof that my memory runs@3600 4 STICKS stable@100%load for days.

Edit: Note that I have heard that the RGB sticks had problems where the ones without were fine, I can't remember where I read that.

Welp, there it is, Flat out 100% CPU utilization 16-16-16-36 1T plus you have a serious OC on.
Can you grab a quick capture of your Wattage utilization? Did you change any of the power settings on your CPU in BIOS? Further, how is your OC enabled? From bios or using Ryzen master?

 

[ub4ty]

Well, I can't screen copy bios, but its the only non-default choice. I pick it, and it knows its 3600 ram, and I leave it at that. Set the vcore to 1.2, and the speed to 3900, and BAM, thats what I am running. Everything else default.

Bios version is 1.3 or 3.0 or something like that. Whatever the newest was 2 weeks ago.

Ah', did you do a bios update Mark after you got your board? Trying everything I can do to avoid RMAing this RAM

 

[ddogg]

What do you guys suggest besides Prime95 and Aida64 for stability testing? I just ran both for about an hour @4.1 without failures at 3443 RAM speeds. Aida64 failed after 30mins at 1.375V so I had to bump it to 1.38.

 

[ajc9988]

Why do you need VRAM blocks for this? I initially played with the idea of watercooling the ram but the budget was already blown out of proportion lol

Not Vram, VRM = voltage regulation module. Even though I haven't had them above 60C so far, and they are rated for higher, I like to keep it as close to 25C as possible. Smoother power delivery and slightly less wattage which comes from heat and electron leakage. It is overkill and a luxury/bling thing, but there is some benefit, even if minimal.

Welp, there it is, Flat out 100% CPU utilization 16-16-16-36 1T plus you have a serious OC on.
Can you grab a quick capture of your Wattage utilization? Did you change any of the power settings on your CPU in BIOS? Further, how is your OC enabled? From bios or using Ryzen master?

Also, if you have geardown mode enabled, even a 2T seeing will read as 1T as it is seen as a1.5T setting.

Ah', did you do a bios update Mark after you got your board? Trying everything I can do to avoid RMAing this RAM

1.30 came out around the 9th or 10th. It is the most current one available.

 

[ddogg]

Not Vram, VRM = voltage regulation module. Even though I haven't had them above 60C so far, and they are rated for higher, I like to keep it as close to 25C as possible. Smoother power delivery and slightly less wattage which comes from heat and electron leakage. It is overkill and a luxury/bling thing, but there is some benefit, even if minimal.Also, if you have geardown mode enabled, even a 2T seeing will read as 1T as it is seen as a1.5T setting. 1.30 came out around the 9th or 10th. It is the most current one available.

Oh VRM, I kept reading that as vram. Ek is coming out with a monoblock for the Zenith that will cover the vrms as well. So if you have a Zenith you're in luck. Not sure about other boards.

 

[Markfw]

What's your temps? I noticed at 3.9@1.2V, my bench scores went down. At 1.225, they scaled about in line with 3.8-3.85 scores. But, I also tested pushing the heat just up to 68C on the die. At that temp, I saw current throttle from 188-220 down to the occasional 145-165W. This explains most likely part of the missing score on those that do scale linearly. Until I have the custom loop in, I can't do my full analysis though. But per 50mhz, you should get an additional 40-65 points in CB15, from what I've seen so far. But I have A LOT more testing to verify this to a certainty.

In that screen copy, 63c, at that temp I have found no problems. I did a CB15 run@4.1, but I am sure it was throttling, but I still beat that 7900X@6 ghz on LN2, so I am happy. I think it was like 3408, but with proper cooling I think I can do 3500 and 4.2

 

[ajc9988]

What do you guys suggest besides Prime95 and Aida64 for stability testing? I just ran both for about an hour @4.1 without failures at 3443 RAM speeds. Aida64 failed after 30mins at 1.375V so I had to bump it to 1.38.

I do a blend. P95 is last. I usually start with wprime loops. If it passes that, then I do burn in test and aida stress test. If those pass, then linpack and OCCT. I then finish with P95. So, it is a slow burn, going from easier testing to the dip testing and aida, to a memory intensive linpack and the cycling on OCCT, then the full burn. I run the later ones longer than the early ones. But that is my way of doing it (I also do Intel stress test, not IBT, before wprime loop on Intel systems).

I also vary which version of p95 depending on whether I need AVX verification at that point in my testing, but always have it for the final test (heat profiles is why).

 

[Markfw]

Ah', did you do a bios update Mark after you got your board? Trying everything I can do to avoid RMAing this RAM

Yes, I think it was 1.0 and I went to 1.3. Is that the RGB ram ? Like I said, that could be the issue.

 

[ub4ty]

Thanks to everyone providing timely help as we try to narrow configs and brings some of our setups to stability.

Also, has anyone with the current issue (not temp related) thought to flip the prochot switch on the mb? I am waiting for the custom loop with VRM blocks to do so, but that may get around the limit you see on current.

Just poked my head in my case and checked. It is flipped [on]. Checked the mobo manual and it says this is the default setting.

I am going to try the bios update to 1.3 (best thing to do to start) and see if this brings stability. Will post updates as I will be trying to bring stability throughout the day.

 

[ajc9988]

Oh VRM, I kept reading that as vram. Ek is coming out with a monoblock for the Zenith that will cover the vrms as well. So if you have a Zenith you're in luck. Not sure about other boards.

EK will have a monoblock for the Zenith, the MSI, and the Gigabyte boards. Not the Asrock boards, which I have the Taichi.

In that screen copy, 63c, at that temp I have found no problems. I did a CB15 run@4.1, but I am sure it was throttling, but I still beat that 7900X@6 ghz on LN2, so I am happy. I think it was like 3408, but with proper cooling I think I can do 3500 and 4.2

Nice. See my post above on HPET on and off on scores. On the win 10 ent. Fall creators update, the real time clock bug is present, although I don't get that warning from Realbench on the spring creators update, so it may just be that version being a beta. Also, I found improved scores with 3333@CL14 than 3600@CL16. Something you might check for your system.