Hey guys, I build a PC with a 2x32GB, 5200 MHz RAM kit with the expectation that I would just add more RAM later… well, now is later, and I’m discovering how messy this is… I don’t need to squeeze out every bit of RAM bandwidth out of my kit, but I need it to at least operate in the ballpark of the advertised speeds.
Has anyone had success running a 5200-spec’d kit (stably) at 4600 MHz?
Have you tried bios 925? any success there?
I only see 922 available for the proart and unfortunately on 922 I had issues booting at all with that bios so it didn’t have a benefit for me
I seem to have been out of the loop too much:
AMD had to address memory issues with Genoa promising fixes, 2DPC doesn’t seem to work reliably. Doesn’t that pretty much sound exactly like the current issues on AM5?
Yes, the MTC20C2085S1EC48BA1R. In Germany there is a pattern of them not being available at all (a single listing with several weeks estimated delivery time, so basically just something to collect preorders), then for a few days a dozen or so shops had them and then they vanish again. Could be the rhythm of container ships arriving.
Supply and demand seem to be extremely out of whack here.
Following buildzoid and kc16 examples I’ve been able to get my Kingston Fury DDR5-5600 CL36 AMD Expo version to run pretty much stable at 5200 in a 4x32 config. It is working 100% stable inside Windows (with many hours of TM5@ABSOLUT, mprime and other benchmarks) but when entering the BIOS sometimes it will freeze, not sure why. I’m running a 7950X on a ASUS ProArt X670E with the iGPU disabled. Any clue about what could be going on with the BIOS?
Are you absolutely sure its a memory issue or might it be your curve optimiser setup is just slightly unstable inside the bios?
Nice memory settings otherwise that is great to see working at 5200.
The curve otpimizer settings were like that for a long time with no problems at all. I optimized them some time ago and I haven’t touched them since then. With this settings I get 39K on CB23 10 min test and the temperature is locked at 88C. The problems with the BIOS started when I went over 4400.
I only have those kits on hand.
Kit 1: Kingston FURY Beast DDR5-5600 CL36-38-38 1.25V (2x32GB) (2-rank, Hynix die)
Kit 2: G.Skill Flare X5 DDR5-5600 CL36-36-36-89 1.20V (2x16GB) (single-rank, Samsung die)
I successfully run those kits together 96GB total at DDR5-5600.
I use these termination settings as below. Also I bump “DRAM VDDQ Voltage” and “CPU VDDIO Voltage” to 1.35V. Rest of timings are just expo timing or AUTO.
Ok, I’ve found something strange here, even with 5200 CL36 stable as in tm5 absolut 10 hours and mprime large FFTs 8 hours. As soon as I run custom FFTs with 224K size it will fail very quickly. I would associate that size to a unstable CPU undervolt, so I removed any undervolt, but still it fails.
At 3600 default voltages it can run 224K even at all cores -15UV with no problem. Also this only happens with PBO2 activated. Using this small FFTs as a super quick test I’ve been unable to go over 4000, as son as I start to increase the voltage over the auto settings to support higher speeds, it will fail in minutes. But if I leave it a 4000 and set the voltage manually to VDD=1.25, VDDQ=VDDIO=1.20 it will work fine.
Running only 2 DIMMS I have no problem running them at 1.3V@6000 with CL36.
Is there anything on the overvolting of the memory that could affect the stability of the CPU? I understand that the IMC is in the CPU so probably that’s the problem. I’ve read somewhere that I should try offset voltage for the VSOC instead of Auto or manual, but I do not know what offset to start with, any clue?
Try to bump VDDP to 1v. By default, it is 0.8v.
You may manually set Vsoc to 1.25v. Below 5600MT/s, Vsoc should be fine in auto as well.
I tried that, but didn’t improve things.
I don’t have 4x32GB yet. I am on 2x32GB+2x16GB for now. I will report back once the 2nd kit arrives. Maybe there are some quirks that I don’t know yet.
I switched to memtest86 instead of testing inside os. Memtest86 is very good at finding error. It stresses both memory and caches. For some pbo settings, you could be running it for half year that seems to be fine but failed in memtest86 with in an hour. I have removed all my overclock settings. You need to make sure your cpu is absolutely stable first, too.
With the termination values and voltages from @kc16 I managed to boot 5600 on Hynix M Die ECC server UDIMMs. All timings are on auto.
I could start tweaking these timings but for now I have to work with this machine before tinkering a little more, so back to more safe settings. But 128GB at reasonable speeds is definitely possible.
Everything on auto won’t even let me boot 4800, so if someone is very knowledgeable I think we could see great results from this platform with 128gb
That’s very interesting! Are you using these Kingston modules? And ECC works also at non-stock speeds? Do any of the BIOS settings get locked if you activate ECC? Sorry if that’s a stupid question but I have no experience running ECC memory.
yes, exactly these ones although not all 4 sticks visually match (had to order through different vendors as none had 4 sticks available)
Yes ECC works at higher speeds too but I didn’t test it at 5600 yet with 4 sticks.
BUT!
These sticks come with 2 expo profiles rated for 4800 and 5600 at 1.25 Volts and with 2 sticks it works without issues. I even managed to overclock to 6000 with CL 32 including enabled ECC when running 64GB only.
No flicking ECC on is just one menu point and you can then decide to enable error injection or not. ECC works fully on this platform at least with the Asus mainboard I have.
Thanks for the info! I was particularly interested as I have the same board and looking at RAM options and ECC seems interesting. I got the cheapest RAM I could find a few months ago with the idea of upgrading when prices on 64GB 6000 MT kits dropped (which they did by a lot).
As the RAM is so finicky I am wondering if ECC would help with stability at all. I would guess that any errors from bad timings or clocking too high would show up in ECC messages? Or does it not work like that?
Yes these errors show up in the event viewer in windows under the WHEA logger. Careful though since injecting these errors can also lead to a system halt to prevent data corruption. I haven’t had enough time to set this up but I bet you could tell Linux and windows how to behave when detecting multi bit errors.
It will make your system a tiny bit slower but RAM that has to correct a lot of errors due to overclocking will spend a lot of time correcting and slow down the system quite noticeable or as said above: Halt it.
I’m also very new to the whole topic but for my work the tiniest amount of instability usually results in crashing software and I’ve never experienced that with Xeon or Epyc workstations and I guess the main reason for that was ECC RAM.
Do you have PBO2 enabled? Can you try to run mprime in custom with size 144 to 224, all avx disabled and using almost all the memory? Doing that will quickly fail in my case, even with a totally TM5-stable 5200MT/s config.
I can try that when I get back home tonight. I’m currently in the process of testing another 128gb configuration. I haven’t enabled pbo but use the 105w ECO mode.
