Instability on idle-to-load transitions (X370 Taichi)

[exscape]

I built a new system with an ASRock X370 Taichi (3.20 BIOS, latest as of today) and a Ryzen 7 1700 this week. It overclocks decently, but as I can't get 3900 MHz to work together with OC:ed RAM (though it's stable with RAM at 2133), I went to 3800 and got it stable at a lower voltage.
It's stable in OCCT, both in "OCCT Small" and Linpack w/ AVX enabled, with the following settings:

CPU: 3800 MHz @ 1.250 V (LLC level 5), tested for many hours, plus gaming
CPU: 3800 MHz @ 1.225 V (LLC level 5), tested for 2h 15m
CPU: 3800 MHz @ 1.200 V (LLC level 5), tested for 1h, plus gaming 2 hours
RAM in all cases: 3200 MHz @ 16-15-15-32-1T, 1.35 V, CPU SoC @ 1.1 V

However, with each setting -- more often so with lower vcore -- I sometimes get a test failure at the very beginning of the OCCT test, when it ramps up the power usage (about 3-6 seconds in). If it passes the initial 20 seconds, it also passes an hour. Running for ~20 seconds, stopping for 5 and repeating seems to trigger the problem relatively quickly; 5-10 tries seems to do it most of the time.

I've read up on Vdroop and LLC and think I have a decent idea (I have a basic understanding of EE concepts) of how it works, and why LLC is considered bad.
However, since I use a low vcore and overshoot of even +0.2 V during load-to-idle transitions wouldn't really harm the CPU, I tested it anyway to see if it would help tackle this problem.
I tested levels 5 (most Vdroop), 3 and 1 (idle voltage = load voltage to within 0.001 V; it never changed a single millivolt during testing idle/load transitions), and managed to get a failure result out of all of them.
The LLC tests were at 1.2 V core voltage.

Any ideas on what I might do to take care of this?
Since it appears stable during load at 1.2 V, it doesn't feel right to go up to 1.275 V 24/7 (if even that is enough, I haven't tried it) just to take care of a problem that probably only lasts for milliseconds at a time!

 

[Reinvented]

Just to burst your bubble a bit, were you testing with a DMM??? Because if you were using software to read voltages, you're going to be in for a huge surprise that they aren't correct or actual. Any sort of overclock, you'll get better stability from using LLC 2 for both. And why are you using such odd timings for memory? What is your kit rated for?

List out your specs, so we can further help you.

 

[BonzaiDuck]

IGNORE THIS POST. I MISREAD OP'S CONFIGURATION AS A COFFEE-LAKE AND INTEL CHIPSET. HE IS USING A RYZEN AMD BOARD. I'll try and figure out how to delete my drivel altogether -- sometime later.
=======================

I wouldn't speak much of my EE understanding, and I'm only seeing your experience through the predecessor Skylake lens. But through at least a few generations since Nehalem or Sandy Bridge, I thought I had figured out a way of reading the voltage impact of LLC settings, and I thought I'd figured out a way to guess-timate the vDroop under these scenarios with monitoring software. I also think I'd managed to estimate the default (Auto) level of LLC for default or stock CPU setting profile.

If you wanted me to describe any of those "methods," I would be hard pressed to recall exactly what I did, and I'd probably spend hours going over my notes of discovery for the i7-6700K and Z170 chipset. Nor would that necessarily avoid confusion in some reference crosswalk to the Z370.

However. Throughout all of this, I simply decided for myself that it was always good to have a residual vDroop, and that you would like to avoid overshoots exceeding the VID used by the processor at certain speeds. And there's been a consistent pattern in my overclock settings, so that I usually choose the second-highest sustainable clock rate, if I can reach it with VCORE recorded at never more than 1.408V [(on my Skylake. Wouldn't know about Coffee Lake, and my guess would be to find out what default VCORE (Auto) is set on the board. Or what Intel likely told their "board partners" would be maximum acceptable Auto VCORE for the processor on a production sample with a range of common defects in an acceptable level of quality.)]

And 1.408 gets me to within 100 Mhz of the binned 4.8 Ghz for my processor. I think I even forgot exactly the value of the stock Turbo for my Skylake. I don't think I ever ran it at it's stock speeds for more than a few hours.

Something else (addendum): Some boards offer superior monitoring. some boards provide little solder-pads in plastic cages for using a multimeter. But with the voltage expectations of these CPUs, you can pretty much tell if your monitoring through BIOS or OS is out of whack. There's enough published data that allows someone to extrapolate a determination of monitoring accuracy.

In earlier chip generations (to my knowledge), idle-level instability might occur so infrequently that it is hard to troubleshoot. Solutions included changes to C-state and reporting in the BIOS. Also, pushing the Offset voltage too high for an overclock result can lead to instability at idle. I had seen advice for as many as three generations of more recent CPUs to try and set the Offset 0 or 0+0.005V -- whatever the motherboard increment. Similarly, there may have been a caution for running a negative Offset value.

If you can do it or grow comfortable doing it, you can tune your CPU speed and voltage with EIST Enabled. You shouldn't need a fixed setting if you can see the various monitored results of voltage states. Using a fixed setting is not going to help you reach a stable overclock more expeditiously. It's apparently just a "mainstream practice."

 

[exscape]

Just to burst your bubble a bit, were you testing with a DMM??? Because if you were using software to read voltages, you're going to be in for a huge surprise that they aren't correct or actual. Any sort of overclock, you'll get better stability from using LLC 2 for both. And why are you using such odd timings for memory? What is your kit rated for?

List out your specs, so we can further help you.

I didn't test with a DMM, no. I probably could, if there's a decent test point reachable by the case's backplate access hole. I suppose I could test stability LLC level 2 as well.

I'm using those timings because it's what I got when I tried to use 15-15-15-32 (tRAS = 32 with the rationale of tRAS = tCL+tRCD+2, which seems to work well).
Those, in turn, I tested as I went down one step at a time, from 18-18-18 through 17-17-17, 16-16-16, 15-15-15 (all of which are stable) and 14-14-14 (doesn't even post).
I read yesterday that Ryzen motherboards don't support odd CL values without Geardown mode enabled, so I suppose that's why I end up with CL16 according to software.

The kit is: Corsair Vengeance LPX Black DDR4 PC28800/3600MHz CL18 2x8GB (CMK16GX4M2B3600C18)
... which I bought as it should be a Samsung B-die kit. The best I managed to run it at (with stock CPU clocks) is 3466 MHz @ 1.35 V with timings of 15-15-15-32-1T (again with the reported CL of 16). 3600 wasn't stable during my initial testing, but I assume that's due to the IMC and not the RAM itself.

 

[Reinvented]

I didn't test with a DMM, no. I probably could, if there's a decent test point reachable by the case's backplate access hole. I suppose I could test stability LLC level 2 as well.

I'm using those timings because it's what I got when I tried to use 15-15-15-32 (tRAS = 32 with the rationale of tRAS = tCL+tRCD+2, which seems to work well).
Those, in turn, I tested as I went down one step at a time, from 18-18-18 through 17-17-17, 16-16-16, 15-15-15 (all of which are stable) and 14-14-14 (doesn't even post).
I read yesterday that Ryzen motherboards don't support odd CL values without Geardown mode enabled, so I suppose that's why I end up with CL16 according to software.

The kit is: Corsair Vengeance LPX Black DDR4 PC28800/3600MHz CL18 2x8GB (CMK16GX4M2B3600C18)
... which I bought as it should be a Samsung B-die kit. The best I managed to run it at (with stock CPU clocks) is 3466 MHz @ 1.35 V with timings of 15-15-15-32-1T (again with the reported CL of 16). 3600 wasn't stable during my initial testing, but I assume that's due to the IMC and not the RAM itself.

What do you mean those are the timings you got? Were you using the A-XMP profile settings? And I am pretty sure those aren't B-die at all, but either Hynix or else E-die. What revision are they?
As far as the memory goes, if you want to overclock your memory with CPU, you MUST increase vsoc voltage to a minimum of 1.1v.

So, since you are probably new to building computers and overclocking let's assume you don't have the right settings. Are you going for all core clock? Or p-state overclocking? Are you using the correct load line calibration to reduce vdroop? What CPU are you using? And not every chip is going to be a winner. Surely, you understand how the lottery works. Sometimes you are going to need more voltage, than your 1.2v that you want. When you test for stability and overclocking, you go for max speed and max voltage, and test. If it passes, then you dial voltage down till it becomes unstable, and then move up another notch. The purpose of this is so you can fine tune voltage control using LLC and offset voltages.

 

[exscape]

What do you mean those are the timings you got? Were you using the A-XMP profile settings? And I am pretty sure those aren't B-die at all, but either Hynix or else E-die. What revision are they?
As far as the memory goes, if you want to overclock your memory with CPU, you MUST increase vsoc voltage to a minimum of 1.1v.

I used the A-XMP settings to begin with, yes, but lowered them step-by-step while testing with memtest86+ and Linpack, until I found the tightest stable settings (those mentioned several times above).
The XMP settings weren't stable at 3600, but were stable at 3466.
How can I check the RAM revision? I was going to check the chips with thaiphoon burner, but Windows Defender gives a trojan warning and I figured it doesn't really matter, so I didn't bother using it. I'm happy with the RAM OC regardless. The CPU stability issue is what matters, and I don't think it's RAM related. (I could test at 2133 with the XMP timings.)

So, since you are probably new to building computers and overclocking let's assume you don't have the right settings.

I started overclocking 1999, so I take that as a bit of an insult.
Though I've been out of the game for a while and mostly used laptops for quite a bit.

Are you going for all core clock? Or p-state overclocking?

I haven't decided, but possibly P-state. Only once I've found fully stable settings to use for at least two states, though!

Are you using the correct load line calibration to reduce vdroop? What CPU are you using?

I mention both these in my first post.

And not every chip is going to be a winner. Surely, you understand how the lottery works. Sometimes you are going to need more voltage, than your 1.2v that you want. When you test for stability and overclocking, you go for max speed and max voltage, and test. If it passes, then you dial voltage down till it becomes unstable, and then move up another notch. The purpose of this is so you can fine tune voltage control using LLC and offset voltages.

Well yes, of course. The thing is, though, that I'm bothered that it's stable during load, but not stable during transitions. It would be a no-brainer to increase vcore if it failed about minutes or even hours of stress testing, but it *only* fails during load shifts.
Increasing voltage 100% of the time to account for problems during the 0.1% of the time where it transitions seems like a waste of energy, plus it adds noise (or shortens CPU lifespan). I was hoping for a more elegant solution that focuses on the issue in a finer manner.

 

[Reinvented]

I used the A-XMP settings to begin with, yes, but lowered them step-by-step while testing with memtest86+ and Linpack, until I found the tightest stable settings (those mentioned several times above).
The XMP settings weren't stable at 3600, but were stable at 3466.
How can I check the RAM revision? I was going to check the chips with thaiphoon burner, but Windows Defender gives a trojan warning and I figured it doesn't really matter, so I didn't bother using it. I'm happy with the RAM OC regardless. The CPU stability issue is what matters, and I don't think it's RAM related. (I could test at 2133 with the XMP timings.)


I started overclocking 1999, so I take that as a bit of an insult.
Though I've been out of the game for a while and mostly used laptops for quite a bit.


I haven't decided, but possibly P-state. Only once I've found fully stable settings to use for at least two states, though!


I mention both these in my first post.


Well yes, of course. The thing is, though, that I'm bothered that it's stable during load, but not stable during transitions. It would be a no-brainer to increase vcore if it failed about minutes or even hours of stress testing, but it *only* fails during load shifts.
Increasing voltage 100% of the time to account for problems during the 0.1% of the time where it transitions seems like a waste of energy, plus it adds noise (or shortens CPU lifespan). I was hoping for a more elegant solution that focuses on the issue in a finer manner.

If you are so concerned about wasting energy, then simply do not overclock. You cannot have it both ways. As for it failing, during transitional load, as you even know that boils down to not enough voltage. So, you either have to pick...higher voltage for stability, or no overclock. There's no "finer manner". It's very black and white here. Even p-state overclocking isn't very useful, because it still draws just as much power. The only thing you'd be doing is reducing heat. Also, you want to use LLC2, period. No if's, and's or but's. You'll find it quickly become more stable.

So simply, up the voltage and enjoy your overclock or don't overclock at all.

As for the memory, yes you can use TB, or you can pull the sticks and read the label. As for testing with Memtest 86, that's a horrible way to gauge whether it's stable or not in this day and age. There's other options that can put a better load on it to test for stability and you should be testing with both CPU and memory at once. Not one or the other. You won't find stability that way, period.

As for the XMP profiles, if that's what it sets, it should be okay. However, you need to up your voltage again to get tighter timings. And usually at 1.4v with VSOC at 1.1v.

 

[exscape]

If you are so concerned about wasting energy, then simply do not overclock. You cannot have it both ways.

Honestly, I don't see why not. Obviously you'll have to live with higher power usage during heavy loads, but vastly increased power usage while idle isn't a must. (The cores use about 10 times as much power idle OC'd compared to stock, since I have a fixed frequency and fixed voltage.)
Besides, the main issue is really temperature and noise during load, rather than +/- 10 watts at the mains outlet while idle.

As for the memory, yes you can use TB, or you can pull the sticks and read the label. As for testing with Memtest 86, that's a horrible way to gauge whether it's stable or not in this day and age. There's other options that can put a better load on it to test for stability and you should be testing with both CPU and memory at once. Not one or the other. You won't find stability that way, period.

That's why I also used Linpack. It uses a lot of RAM, but I don't how it uses it, so it might not be very stressing for the RAM. I also don't know how well it tests for errors.
What software would you recommend?