So I wanted to build a replacement in house server for some light duty OPNsense on a VM and another VM to host games. I looked at threadripper and epic, but, they cost way more than I was willing to spend. Instead, after watching a level one video on am5 48GB dims, I thought I would give 4x48 a shot on am5 as a replacement platform. It should be much better than the old i7 6850k at 4.2gh all core(6 core). Its been solid but its time to update to this decade.
I finished building tonight and it seem to be ok. Several bios flashes later, it looks like 1003 is a winner for asus boards as long as you dont have an x3d chip. I have a 7950x so I thought why not. I had read in other forums that this one worked for the asus pro art x670-E No windows yet because I would like to really hammer this new build with some memory tests until I can be sure its stable enough to be reliable. I know it will not be as good as it could with ecc memory but I think that might be more than I really need.
Does anyone have any tips or tricks for how to add stability or to test for stability? It seems to be stable at the corsairs xmp profile 5200 but if I can test for and detect any instability I will want to play around with SOC, Mem voltage and maybe a few other things until I can make it as stable as it can be on this older bios.
Check the QVL list, if it is on the QVL list. I wonât bother to stress test it.
There should be a particular version of the bios stated, âAdded 24/48GB DIMM supportâ. Used that one or any version after that.
Bootable memtest 86+ to start (so there is no risk of curruption if you crash), then TM5 if youâre on windows. Computational stress tests like p95 (with large FFT size) or ycruncher are also good.
I have been running memtest86+ 6.2 on the pc for 7hours now with all 10 tests selected and it still says pass and has completed 2 full runs. It looks to take a little over 3 hours to do a full pass of all tests. I will let it continue to run for a while but I am unsure how long.
Sometime today, if it is still passing, I will stop the test and install the OS. If I was hardcore I would let this run for a few days before I determine it works âwell enoughâ but its not like I am hosting anything really mission critical. This servers primary use is âfor the lolsâ
I agree buying something on the QVM list is a great idea, assuming you can find what you need. I did look at the qvm memory list for this MB, but, i did not see anything close to what I wanted(192GB with RGB for the pretty pretty lols).
Great suggestions here! If anyone has any questions or more suggestions, please post what your thinking and maybe more folks can get 192GB working on their own rigs.
It should also be noted that I have a 40mmx20mm noctua fan sitting on top of the memory clips. In am4 with 4 dims this was needed to keep the memory stable. It was fine with the leds turned off but when the led were set to full white they would creep up to mid to high 60s and get a little unstable. The fan keeps them below 60 always and they are solid. It might be worth a try if you have 4 dims and any memory temp issues like I did.
Follow up - I had to play with the soc, memory voltages, and termination values with the latest bios, but, it seems rock solid now. I will post my values for this MB and Ram in case that is helpful for others.
Its real picky but once you find the right combo it works great.
The defaults for expo RZQs are 5 5 2 6 6 just like in the picture above showing 5 5 1 6 6 didnt really need any changes from expo. These are prett solid and nothing was ever better than these except for one bump either way on RttWR. Lower seemed better. default(two) is probibly safer for most. Expo defaults for the win here!
The proc ODT only had good results between 48 53 and 60 and only at some voltages of memory and SOC. In retrospect these did not need any changed from the xmp so I would not bother unless you really need to try something different. if default(48) works thats probibly ideal
CLDO VDPP This MUST be lowered to post reliably. .9 works fine as does anything lower. I think .98 also works fine but default is like 1.15. Not sure what this is for and it may be related to onboard graphics??
I set all the power phase profile stuff to âOptomizedâ and 130%
On board graphics disabled. I think its fine enabled but i did most of my testing with it disabled.
SOC defaults to 1.22 which is fine, mostly I think. If you can get away with 1.12 or something like that I hear that is better but I always had more stability around 1.25ish. I think Load line calibration may make it slightly more or less stable. With SOC load line either go 1, auto, or like 4-5 and see which one is better for you
VDD and VDDQ both needed a slight bump to stay strong .26 may work but .27 just seemed better in most tests. It would post and run strong at 1.35 1.36 and may go up to there in one or both. Less is more here unless you really need it.
I tested over a week to probe all the setting and learned a lot about this combo. Sadly I would not have need to test so much if I had figured out that the graphics card and nvme card were being pulled to hard by the mounting bolts making the system intermitantly unstable. The worst kind of unstable! It acted a lot like power or ram issues. The first time I solved a problem by remounting the graphics card was a clue. The second time was when I realized what was happening and as soon as I bent the back of the case like 1.5 mm closer, the bolt holes for the cards lined up and no more pull and no more issues! DOH, what a rookie mistake!
Here is a pic of me stress testing âMissMinutesâ. I have a disk benchmark running, and Prime 95 running, and cinibench, and furmark, and to really add to the punishment, I also made MissMinutes play Mileyâs wrecking ball on a loop! Take that; sanity!!!
Terminations, if leave all auto, sometimes it will screw up RttNomWr and RttNomRd, which gets stuck at memory training. I just hard code these two to 48, and leave rest of them auto. It works just fine.
VDDP, it wonât boot with 1.15v. The sweet spot seems to depend on VDDIO. If I use VDD=VDDQ=VDDIO=1.25v, the VDDP sweet spot is at 9.2v for me. If I pump these 3 memory voltages all to 1.35v, the sweet spot also moves up. I donât really need 1.35v memory voltage, as any frequency above 5600 is never stable for me.
Soc voltage, if you run fclk 2000, you probably need to use at least 1.1v Soc. I use 1.15v it should be good up to 6000.
VDDQ and VDDIO also have a sweet spot. 1.27v is also what I found. As said, it also depends on VDDP. Keeping 0.3v delta between VDDIO and VDDP seems to be optimal.
Nice!!! Once I have tested the final config a bunch and if it looks good I may try the settings as you suggest and see how it impacts heat and performance. Maybe I could even get a faster speed working. I have heard its best to use the lowest SOC that is solid. Great info on the VDDP, thank you!
Some interesting testing in the last few days. Sometimes I would boot and be able to run memory tests and furmark all night without an issue and othertimes it wouls imediatly shutdown on the first attemmpt or even loading into windows if I have Hyper-V setup.
After inspecting the settings I needed to set ProcDQDS to 34.3 and RTTWR to two and back to stability. Then it would do the same thing repetably until I changed thoes to ProcDQDS 40 and RTTWR to 1. Back and fourth it would go on some reboots.
Tried a bunch of setting to stablize this and not sure if I have found it but MISC Voltage was defaultig to 1.128 so I bumped it to 1.2. Really its set to 1.185 but reads as just over 1.2 in the bios when set to manual. I also set ProcDQDS back to 40 and RTTWR back to 1 as they are very responive, when working, even under heavy load. I also turned off LPWR.
I will test more today and this week to see if I can sort this out. I know there is a bios setting to switch profies on the fly so single threaded and miltithreaded can work at their best. Maybe this has something to do with the issues?
Stay tuned for more fun in testing results. Of course I could just buy supported memory in the supported ammounts butâŚwhere is the fun in that?
Any other suggestions to address this are most welcome.
last time i tried setting up hyper-v it just fell over. repeatedly.
ended up switching to virtual-box instead⌠yeah its not as integrated, but at least it works
Did you try 5600 or 6000? I saw some screenshots with 192GB running at 6000.
As you found, there are huge memory training inconsistencies. You changed some settings, and it worked, which maybe just due to the inconsistencies. It not necessarily means better settings.
Its 5200 ram. I could boot at 5400 but with lots of errors so I did not play with that really at all. 5600 and above wont boot.
It looks like it was vpp voltage needed to be 1.9v. Its hard to say for sure but once I set that it stopped acting up.
To test I rebooted and ran all the tests at once. I did this 10 times without a hitch. So I thought if its really VPP then it should go back to wonky at 1.8. To test that, I rebooted and tried to break it 10 more times but it stayed stubbonly solid.
I will let it run overnight again tonight. Maybey the vpp gave it the bump it needed to train properly? I dont know for sure but I cant get it to crash or act wonky no matter what I thow at it.
I will keep testing intermitantly but it seems solid now so I left VPP at 1.9. My updated settings are listed below with the non defaults marked in yellow.
