TL;DR:
maxed out negative offset on CCD1 and basically -30 on CCD0.
Despite all evidence to the contrary, that cannot be stable right?
I must be am missing the right stresstest?
———–
OK, last post got me no comment on RAM so I guess I will lock that in at DDR6000, TCL 30, FLCK 2200. Now I am back again after a surprisingly short run of curve optimizer testing. I followed this guide with a few tweaks. https://www.youtube.com/watch?v=rY2nlI84S0E
Reminder, CPU is a delidded 9950x3d on a custom loop.
I did not set motherboard limits or 10x scaling. I stuck to PBO enhanced with the level 2 thermal limit (80 degrees) and with an x3 scalar. Likewise i went with only a 100Mhz as per his recommendations. I also did not use curve shaper.
I ended up with a CO that seems too good to be true?
I completely maxed out CCD1 at -50. i didn't try going more as this was the maximum for the SMUDebgug tool and I did not expect it to be actually stable despite not crashing. On CCD0 I basically went -30 on all cores. I tried -34 and 36, both resulted in a freeze. So I just backed it up to -30, the baseline offsets not withstanding.
So I ran an overnight corecylcer test with VT3, 10 hours, passed. I also ran an hour long test of corecycler prime95 huge FFTs. An OCCT cpu+ram test for an hour, 30 iterations of linpack extreme. All run fine, no WHEA errors, no clock stretching. All of this with no curveshaper applied and global c-states enbabled.
I then played several hours of darktide at max settings, 5120×1440, raytracing enabled. The HWinfo screens are based on darktide, must have been 3-4 hours or so. No issue, along with a timespy stresstest. The only thing that failed was AIDA64 with fpu and cache selected alongside CPU. I dismissed that though as it fails immediately after starting and passes if I leave only CPU.
Seemed a little suspiciously sensible compared to the other results. R23 also passes.
You can see from the screens the CPU sits around 50ish degrees during darktide. That was why I didn't bother with curveshaper. Even with a somewhat demanding load, both light and heavy loads would all fall in the same frequency/temp band. That should make it unstable from what I read, but that doesn't seem to be the case?
So not only do I find myself with a CO offset that from what I can tell is unheard of, on top of that it works without adjusting curveshaper? There is just no way, right? Is there some other test I should run?
If not, I guess GPU overclocking would be next.