Overclock.net banner
1 - 9 of 9 Posts

·
Registered
Joined
·
5 Posts
Discussion Starter · #1 · (Edited)
Its been about 7 years since I last built a PC. Never bothered to OC it. Never ran prime95 on it, obviously. Finally put together a new one beginning of last month. Asrock B550 Steel Legend, 5900x, 32GB G.Skill Ripjaws V (dual rank, F4-3600C16D-32GVKC), 6700XT, 850w PSU. All new.

Running Linux, I left the AMD overclocking section alone for most of the month since I don't trust myself with the lack of Linux monitoring tools for Ryzen. Just set XMP, turned on SAM, set fan curves and all that. No unexpected errors or crashes.

Around the end of the month, I decided I would see what I could do with curve optimizer and dusted off my old SSD and popped Windows 10 on it. I don't have much of any experience overclocking but know my way around computers. I've been playing around at this for almost a week now.

It's already somewhat hazy, but for the most part I had PBO set to manual with the stock values set (142/95/140). I tried higher values at one point (including the pointlessly high motherboard limits to see what the max actual values were) with Ryzen Master in Cinebench a couple times but only ever just barely went over 80c. I may have left scalar on auto for a while but eventually set it to 1x. Checking hwinfo64's report with auto always said it was 1x at any level of load though.

I tried putting a little boost override in at one point of curve optimizing but couldn't tell if it was helping or not so I turned it back off.

CPU/SoC LLC was set to auto for the whole month. CPU on auto is Level 3, which is is the medium setting. 1 being a flat line, 5 drooping the most. SoC auto is Level 3 without XMP enabled and Level 1 with XMP enabled. I've left that one alone.

I started testing with corecycler's default Prime95 config. I started with the method of starting at -5 and working my way higher. I found almost immediately that my preferred cores on CCD1, 0 (first) and 2 (second), didn't like any negative offset. Not even -1. But even at zero offset, cores 2 and/or 0 would rarely come back with rounding errors. And even with the PBO page disabled completely. The preferred cores on CCD2 in Ryzen Master, 8 and 11, seem fine even at -20 and -15 respectively.

I used some corecycler Y-Cruncher, as well as Y-Cruncher on its own with the full list of tests a few times for all core until I noticed an instance of core 4 apparently spiking to 89c when it switched to the small in-cache FFT test. Core 4 Power maximum was 22W in that instance. 87c/20w in the first instance I noticed it. Most of the others in those multicore tests had a max of 9W with a couple 10Ws and 15Ws. Core 2 spiked second highest in those. Almost 10c over the rest which were around 65~70c maximum. But all cores would be 50~60c normally.

Besides those I would occasionally also do some OCCT at the start as well.

At some point I saw that 720-720 FFT SSE is pretty quick at finding errors. With or without XMP enabled and core optimizer disabled, core 2 pretty much always hit a rounding error on the first iteration even at auto (30 seconds, roughly).

Reset BIOS to stock by clearing the CMOS, only turned off CSM to enable secure boot. The test I did was a 5 iteration of 720-720 FFT SSE cycling through all cores. No error. Not very thorough, but after many so many immediate rounding errors I moved on to turning XMP on.

XMP on, 720-720 FFT SSE auto 5 iteration on just cores 0 and 2. No rounding errors. Restarted to double check what some auto settings were in BIOS but didn't touch anything. Tried 720-720 FFT SSE again but cycling all cores. This time core 2 hit a rounding error on iteration 3.

Turned XMP back off, ran another 720-720 FFT SSE. Core 2 rounding error on iteration 1.

I've been keeping hwinfo64 sensors open whenever I test. Highest any core gets with corecycler prime95 with any settings is 75c. Multicore stuff usually tops out in the mid 60s. The highest CPU Core Voltage (SVI2 TFN) has hit was around 1.487~V with CPU LLC3. Can't remember what it was with CPU LLC2 but never over 1.5V. SoC Voltage (SVI2 TFN) is 1.087V to 1.094V at SoC LLC1. SoC LLC3's values dip under 1v.

What are my options here? I don't have any other Ryzen CPUs, DDR4 kits, or AM4 motherboards to test with. I have yet to see any core other than 0 or 2 hit a rounding error at stock settings. I haven't done an extensive memory test, but I did several passes of memtest86 after I got this running that were clean.

EDIT: It seems to be a B2 if that matters.
Rectangle Font Material property Parallel Slope
 

·
Iconoclast
Joined
·
32,534 Posts
Preferred/best cores that can't take negative COs aren't particularly uncommon as they are already the the ones with the most aggressive performance values at stock. If you don't get rounding errors with PBO disabled and the memory at JEDEC settings, it's not technically defective.

In general, on ASRock boards, LLC should be set to 2 or 3 for the CPU and 2 for the SoC.

XMP is a crapshoot and should be avoided. Manually set your timings and voltages. vSoC may need a slight bump, and VDDGs may need to be hand tuned to stabilize any memory/FCLK significantly above 1600MHz.
 
  • Rep+
Reactions: Luggage

·
Registered
Joined
·
5 Posts
Discussion Starter · #3 ·
Preferred/best cores that can't take negative COs aren't particularly uncommon as they are already the the ones with the most aggressive performance values at stock.
So I guess I made a mountain out of a molehill for myself?
If you don't get rounding errors with PBO disabled and the memory at JEDEC settings, it's not technically defective.
Well that's the thing, is that I was getting them with PBO disabled and memory at default. But only rarely on those two cores specifically on small FFTs, and quickly at that 720-720 FFT. I was under the assumption that small FFTs were done entirely on CPU, so memory wouldn't really come into play.
In general, on ASRock boards, LLC should be set to 2 or 3 for the CPU and 2 for the SoC.
I'll try testing CPU LLC 2 and SoC 2 tomorrow.
XMP is a crapshoot and should be avoided. Manually set your timings and voltages. vSoC may need a slight bump, and VDDGs may need to be hand tuned to stabilize any memory/FCLK significantly above 1600MHz.
I guess I should have mentioned what clock the memory was at too. 2133 at default, 3600 with XMP. 1.20V default 1.350V XMP. So 1800MHz FCLK at XMP settings. Not sure where to start on setting timings myself, or what voltage ranges are acceptable.
Trying to set the XMP timings manually won't post with error code 22. Only thing that stands out to me immediately is "tCKE" being set to 0 with XMP, where the motherboard defines the range as 1-31 and 0=auto.
I did do another 4 pass of memtest86 last night on XMP, and it still didn't find any errors.
 

·
Registered
Joined
·
1,894 Posts
OP one of my best cores is at -1

Product Rectangle Font Screenshot Technology


It's completely normal for that to happen (much lower than the rest), especially on CCD1, which often sees the absolute best core of the whole chip.

Cores failing with PBO off, well, before thinking it's a faulty chip, the rest of your motherboard voltages are important to look at. Some motherboards put absolutely stupid voltages in play if left on AUTO.

It's the same with memory as well tbh, XMP only really handles the primary timings and VDIMM. Some motherboards can put VSOC/VDDG/VDDP at crazy voltages, not to mention introduce some suspect secondary timings.

Point being, memory is heavily tied up with the CPU, and instability there can sometimes trip CPU stability testing apps.

Don't mess around with LLC, just not necessary for normal use. It's unlikely your motherboard will do something stupid here.

tldr; AUTO can often minimise user input mistakes/unstable values, but sometimes AUTO is a bit of a joke between manufacturers.
 

·
Registered
Joined
·
5 Posts
Discussion Starter · #5 ·
Yeah, got it. Feels bad that both preferred cores cut it so close but I guess that's what you get with the CPU that's two 5600x's glued together.
Really not sure what voltages I should be putting in myself, or what values I should be even looking for. I thought manually setting voltage was for manual OCs.
Guess I need to find a primer on tuning DDR4 as well.
 

·
Registered
Joined
·
1,894 Posts
Yeah, got it. Feels bad that both preferred cores cut it so close but I guess that's what you get with the CPU that's two 5600x's glued together.
Really not sure what voltages I should be putting in myself, or what values I should be even looking for. I thought manually setting voltage was for manual OCs.
Guess I need to find a primer on tuning DDR4 as well.
The other way to think about it is the best cores are already pretty tuned to boost to the max possible, therefore, their voltage curve is likely to be more optimised than the others.

I think there is just too much "advice" out there that unless you have "THE MOST NEGATIVE CURVE EVER", your CPU is a dud or not performing well. Just not true.

And to be honest, the YouTubers and commentators that lie to themselves that they are -30 all core stable, just end up with reboots, clock stretching and ironically, less performance than anyone who understands what the curve actually does and tweaks it properly.
 

·
Registered
Joined
·
1,440 Posts
The other way to think about it is the best cores are already pretty tuned to boost to the max possible, therefore, their voltage curve is likely to be more optimised than the others.

I think there is just too much "advice" out there that unless you have "THE MOST NEGATIVE CURVE EVER", your CPU is a dud or not performing well. Just not true.

And to be honest, the YouTubers and commentators that lie to themselves that they are -30 all core stable, just end up with reboots, clock stretching and ironically, less performance than anyone who understands what the curve actually does and tweaks it properly.
IAWTP, anybody who claims -30 +200 in PBO for all cores fully stable, is being dishonest.
 

·
Registered
Joined
·
776 Posts
IMHO **** pbo. unless your willing to really tinker pbo is a **** show. do an all core clock that voltages and temps are acceptable to you. that or just leave it stock. the one build i did with a 5900x boosted to 4.9ghz stock we did use a 360rad on it so temps were pretty decent but under single core boost temps would often hit 85~88c all core stock was around 4.4 and tbh thats really OK as your memory latency is going to affect your playability more than the core speed will. core speed will help with the max fps and help carry the lows a bit, kinda doubt they will have much effect how smooth your game is beyond 4.4~4.5ghz even the 5800x3d does not seem to scale with core speed in games. and that cpu is not dependent on memory hardly at all.

i would say a 4.7ghz all core would be the easy choice as most games are using more than 2 cores you never going to see more than 4.7ghz in most games anyway. and in games that use 4 or more cores the 4.7 will be a large boost over the stock operation with less deviation.
 

·
Registered
Joined
·
5 Posts
Discussion Starter · #9 · (Edited)
Well I tried CPU LLC 2 with SoC LLC 1 and 2 but it didn't help. Still got rounding errors pretty quick with 720-720 FFTs in corecycler. When I say that, I mean with suspendPeriodically=1. Is that even a realistic target of stability? Did I misinform myself by a few posts I found somewhere? But even then, regular Small FFT SSE still fails on those cores the same in corecycler eventually.
I'm still to nervous to set any manual CPU/SoC voltages. I guess my next step will be figuring out timings for my memory. This is what ZenTimings has for my XMP.
Rectangle Font Screenshot Technology Parallel

At this point I'm just bothered that the problem exists, even though I haven't had any issues with normal programs.
After all this I mostly lost my drive to keep going with curve optimizer but if all else fails adding positive offset to those cores would probably do the trick?

EDIT: Oh, I just remembered something. hwinfo64's Power Reporting Deviation stabalizes at 108% on CPU LLC3 when running Cinebench 23 multicore. 106% at CPU LLC2. Could that have something to do with it? Over 100% means the CPU is drawing less power than it could right?
 
1 - 9 of 9 Posts
Top