Feedback for 9000 X3d chips

[dochodges]

Rans the automatic config and found -50 was stable for all cores. I know this is not the case so I imagine that the forms of stress test needs tweaking for the auto test with 9000 CPUs

[risingstarfish]

This also happens to me. I'm using run core cycles.bat with my config set to Aida or cruncher.

I think I've figured out a little of what might be going on with these CPUs even though I don't have a 3D L3 model.

The two CCDs are not matched. At all. I don't just mean the L3. It almost feels like AMD's closest equivalent to economy cores even might be going on here. On my 9900X I was able to get CCD1 all cores at -50 with 100% stability under full load while CCD2 had three cores that were quite bad (two at -22 at best and even then probably needed a bit more leeway) but no matter what I did, I'd still get freezes during low usage/near idle. Looking at voltages, I realized those -50s weren't really doing anything as nearly as I could tell. That is to say that the voltage wasn't really going lower than what the best CCD0 cores wanted when under load. Also, CCD1's cores all seem to maintain a lower maximum frequency. (-250MHz I think, but most of the stuff I can most easily look in really likes rounding badly, so it could be -200.) In fact, if I set a negative boost limit to the CPU, CCD1 goes down by that same amount to continue to run lower than CCD0. (Ex if I set -250 so CCD0 is limited to 5.4, CCD1 runs at 5.1 max.) It almost feels like CCD1 was designed to be more efficient than CCD0, like it's for a laptop or something.

I don't think the biggest problem is actually CCD1 exactly, I think it's the mismatching causing the issue. CCD0 generally wants more voltage and while the SOC is trying its best to maintain a weakest link voltage to the CPU, CCD1 using such lower values gets harder to keep up with, especially during high variance (such as not actually under a real load where cores are parking and going up and down a lot.) We have to remember that on stock, all cores are at +/- 0 expecting the same voltage curves, so it maybe just wasn't designed to deal with such strongly conflicting curves when an entire chiplet will gleefully accept so much less while another wants so much more.

Figuring out working values with a mismatch like this is a... tricky process. I'm not 100% sure if the SOC can keep up with maintaining them separately like this under all situations with them so different. I got sick of fighting with it and went with the weakest link, limiting all cores to what the worst core wanted (plus a slight margin.) If I'm right though, I think CoreCycler is going to have to straight up ignore results from CCD1 during all initial testing portions and focus on CCD0, then just find out if any cores on CCD1 want anything higher than the best on CCD0 (which might possibly be unlikely.)

[kerriganx]

Do not set -50 as starting point, you will get freezes even if it looks stable under load.

As a starting point for Ryzen systems you could set the Curve Optimizer to e.g. -15 or -20 for each core and then wait
and see which core runs through fine and which throws an error. Then you could increase the setting for those that have
thrown an error by e.g. 2 or 3 points (e.g. from -15 to -13) and decrease those that were fine by 2 or 3 further into
the negative (-15 to -17). Once you've crossed a certain point however there is no way around modifying the value by a
single point up/down and letting the script run for a long time to find the very last instabilities.

Make sure to add Curve Shaper as well to help with idle freezes.

My settings (9950X3D):

stressTestProgram = y-cruncher
coreTestOrder = CorePairs
mode = "24-ZN5 ~ Komari"
tests = VT3

Curve Shaper min/low to Disabled, med/high/max to -15. Curve Optimizer stating point to -20. No Scalar, no Fmax Boost Override.

You will still get freezes sometime, if that happens just increase values on the cores you haven't touched yet. It becomes a much easier once you found "bad" cores on CCD1. They really ruin everything. Then you can try to lower values for "good" cores.

After endless hours of testing I get following values:

CCD0: -29 -30 -30 -30 -30 -30 -30 -30
CCD1: -30 -11  -4 -22 -21 -22 -30 -29

Curve Shaper med/high/max: -15

As you can see, I have pretty low voltages on CCD0: -30 CO plus -15 CS = -45. But look at CCD1: core 9 with -11 CO, core 10 with -4 CO etc. Those bad cores are the cause of most freezes.
When you're done, add more tests and let the script run for a long time to find the very last instabilities. VT3 seems to detect problems the fastest. But make sure to try the other tests just in case.

[reisholmes]

Just came to say this appears to be happening to me as well but I'm on a 9800x3D so only one CCD.
I ran the 'Ryzen.AutomaticTestMode.yCruncher.ini' settings on defaults with the 0.11.0.0alpha4 release and it passed fine with -50. I personally know from testing before this that -30 all core is fine but -35 will BSOD windows or crash less than 10 minutes in AIDA64 stress test.