Originally Posted by moorhen2
You don't state what Ram kit you are using, you need to run a memory specific stability test, ie GSAT. HCI or Ramtest which I use and tests the cache as well. Also make sure you have the Mem OK Switch in the Disabled position on the mobo.
Well the kit shouldn't be relevant in this scenario as I firmly believe the issue lies in the motherboard and how it trains. I've tested two different kits, 4000 CL19 and 3600 CL16 (G.Skill). I'm currently running G.Skill F4-360016D-32GTZN. Mem OK is disabled, MRC Fast Boot also disabled (to ensure training). I did also state that I already confirmed GSAT stability (which, as I've understood it, primarily tests the stability of dram voltage and timings). The issue lies in VCCIO and training (Karhu and Prime95 112k in-place non-avx both test this and confirm instability in my case, due to bad training). I'm sure the XI Apex does exceptionally well with 2x8GB sticks but dual rank is very, very different. There's a lot of stuff the motherboard does on auto-training that has to be manually tweaked to get it to post properly at high frequencies (generally setting things to tight).
Originally Posted by Imprezzion
Hol' up. Your tRCD and tRP are lower than your CAS? Well, there's probably your problem lol.
Try 4000Mhz 15-16-16-31-350-CR2 or 15-17-17 if it's immediately unstable. Use... I'd estimate... 1.225v SA and 1.175v IO and 1.45-1.50v DRAM? That would be a clean starting point.
Training memory is a difficult process that I don't fully understand but understand enough to know and also experience in my own process of tweaking RTL/IO that everything affects it. Not just one specific voltage or timing but rather a combination of everything together.
For example with tWR 14 tCWL 16 tWRPRE 34 my RAM will not properly train. RTL is off by at least 5 or 6. Setting it even lower at for example tWR 12 tCWL 14 tWRPRE 30 it trains fine. My RAM also tends to fail completely in locking / training if tRDWR or tWRRD is set too tight. It will run "stable" and not error per sè but have terrible performance and missed training.
It looks weird sure but it's not uncommon for B-die and there's no issue with it - I've seen plenty similar cases, like I said all the timings in the picture are 1 hour GSAT stable. The same issue persists even with 18-18-18-40 and very safe, stable secondary timings. 4000MHz is very, very demanding even with a really good IMC and motherboard. Buildzoid has a video of him overclocking a pair of 16GB Corsair Dominator sticks where he mentions only one of his motherboards actually capable of pushing it above 3866, an MSI ITX board that just handles training well.
The kit I have is capable of a lot tighter timings than what the pictures show but set most things a lot safer just to find VCCIO stability first, hoping that'd alleviate any difficulty training.