Finally: Overclocking EPYC Rome ES

[RageBone]

you have the jumper JGP1 on the board that should allow you to disable the BMC GPU.
That might be worth a try.
In the past on other SM boards, the bios setting for Offboard GPU worked as expected.

Is it the modded bios with the AMD CBS options?
Maybe you chose a setting in there instead of the usual one?

 

[dragonian]

I'm beginning to believe though that one (if not both) of the CPUs have a problem with the PCIe lanes.

I have 2x ZS1711E3VIVG5 in 2 different H11SSL-i boards. One of them I cannot use the internal m.2 slot, and a couple of the pci-e slots. (the other board everything works fine) I did some troubleshooting back in the day, and swapped mobos, and re-torqued the cpu. No dice. The best I can figure, is that one CPU just has faulty pci-e lanes. I have accepted this as the drawback of using ES parts.

My only suggestion is to try different slots.

Edit: I solidified this conclusion after trying a nvme drive in a pcie expander card.. It detects fine in some slots, nothing in others. it might be easier to test specific lanes/slots.

 

[grsychckn]

Okay, I tried with two different GPUs and both failed. Interestingly, the V5900 (no external power) wouldn't give me signal, while the Vega 56 behaved identical to the 6900XT (video, but only through POST).

I swapped the Rome CPUs for a pair of Naples 2S1905A4VIHF4 CPUs and gave it another try. It wouldn't recognize the UEFI Windows boot partition so I tried to reinstall Windows. It booted from the USB device then once it entered extended mode, it reboots. I know these CPUs worked before but I've forgotten all the settings and it was on a v1.0 version of the H11DSi. I'm going to put the original Rome CPUs back in but this time I'm going to swap them in their sockets.

 

[grsychckn]

you have the jumper JGP1 on the board that should allow you to disable the BMC GPU.
That might be worth a try.
In the past on other SM boards, the bios setting for Offboard GPU worked as expected.

Is it the modded bios with the AMD CBS options?
Maybe you chose a setting in there instead of the usual one?

The BIOS I'm using for the Rome CPUs is the H11DSi-NT v2.0 one from the OP. Was looking for any suggestions on settings the might effect the PCIe functionality.

 

[grsychckn]

Well, I finally got the PCIe GPU working after swapping the sockets but I had to also disable SVM. I don't think I disabled that feature in the BIOS before, but I got desperate and started disabling one thing at a time. I'm now going take a screenshot of my settings and save them so I don't forget how to get these things running. Thanks all for the help!

 

[Wasmachineman_NL]

AMD "Milan-X" 7773X EPYC CPUs with 3D V-Cache get overclocked to 4.8 GHz - VideoCardz.com

Dual EPYC 7773X with 1536 MB of 3D V-Cache tested and… overclocked Chinese Bilibili content creator ‘kenaide‘ goes through the overclocking process of AMD’s newest EPYC CPU. AMD EPYC 7773X Milan-X CPUs, Source: Bilibili/kenaide And yes, you read that right kenaide took a data-center EPYC 7773X...

videocardz.com

ABSOLUTE MADMAN

 

[mirrormax]

AMD "Milan-X" 7773X EPYC CPUs with 3D V-Cache get overclocked to 4.8 GHz - VideoCardz.com

Dual EPYC 7773X with 1536 MB of 3D V-Cache tested and… overclocked Chinese Bilibili content creator ‘kenaide‘ goes through the overclocking process of AMD’s newest EPYC CPU. AMD EPYC 7773X Milan-X CPUs, Source: Bilibili/kenaide And yes, you read that right kenaide took a data-center EPYC 7773X...

videocardz.com

ABSOLUTE MADMAN

Huh looks like he didn't actually oc
it. He only posted a pretty low r23 score. He 100% did not run it at 4.8.
No server motherboards are made for those kinds of speeds and power.

 

[Wasmachineman_NL]

Huh looks like he didn't actually oc
it. He only posted a pretty low r23 score. He 100% did not run it at 4.8.
No server motherboards are made for those kinds of speeds and power.

yeah they appended the article with input from our @ExecutableFix

 

[FirstEver]

CPU-Z bug (maybe can't support 256-threads), or Threadripper PRO 3995X is faster than this EPYC ROME ES?

OC tool v 1.0.0 works on TR PRO and the latest Windows 11 Enterprise Insider Preview build...

 

[mirrormax]

cpu z is terrible for cpu benchmarks especially high core count. it often dosnt use all threads, open taskmgr while it runs. one run it uses 60% threads, then you restart it and it uses 80% etc.
seems to change from version to version aswell. i found one version that used all threads and managed to get 50k score there but just in general its not great past 128 threads

 

[FirstEver]

I think that the same (better) performance can be achieved (for 64-core / 128-thread AMD CPU's) with the appropriate settings in the BIOS on the ASUS Pro WS WRX80E-SAGE SE WIFI motherboard.

R15


R20


R23


I use consumer DDR4-3200 CL16 (8x32GB).
It works stably 24h / 7days, but ASUS still has a lot to improve on DDR4 (non-ECC) memory support.

AIDA64


Do not ask

 

[mirrormax]

If your original question was if 3995wx is faster than a single 7742 epyc es or retail, then yes it's the same chip but clocked higher and higher tdp so performs better.

But in any program that can utilize 256 threads a dual socket epyc will obviously outperform it by a lot, for example 100k is doable in xmrig, and also cinebench r23

 

[heavyarms2112]

Anyone knows what VRMs are used on Asrock epycd8-2t motherboard?
I was able to use the Rome OC software on an AMD EPYC 7742 QS and applying 4 GHz max core frequency yield ~23.8K pts on CB20 but the power consumption at wall was around 400W
considering a gpu (20w) and chipset power/psu efficiency loss etc. that might still be over 320w. Is the VRM capable of handling this power?

EDIT: For sake of comparison. stock power consumption is 320w at wall.

 

[mirrormax]

I've seen 1000w out of the wall with 2 so should be fine with adequate vrm cooling. Your actual clock was most likely much lower though the oc tool combined with EDC bug can display some wierd numbers, sounds like around 3.2ghz

 

[heavyarms2112]

I've seen 1000w out of the wall with 2 so should be fine with adequate vrm cooling. Your actual clock was most likely much lower though the oc tool combined with EDC bug can display some wierd numbers, sounds like around 3.2ghz

Yup. Actuals were a lot lower. 3.1 Ghz or so. the EDC bug killing the chip doesn't impact 100-000000053-04 or is it no longer relevant?

1000w on dual socket board?
Typically, on these high power CPU boards we got a chunky VRM heatsink. I wasn't able to see anything like that on epycd8-2t board.

 

[shpitz461]

Those of you with an ASRock ROMED8-2T, what settings did you change in BIOS for performance improvements?

 

[mirrormax]

Did anyone test this yet on Milan es/qs CPUs that are now more commonly sold?

 

[bayleyw]

Now that this tool has been around for a couple years, I'm curious what volts folks are running 24/7 on their CPUs. TSMC 7nm has always been a little iffy in terms of max safe voltages.

 

[RolloZ170]

TSMC 7nm has always been a little iffy in terms of max safe voltages.

1,5V over 6month was tested, nothing happened. i think its more the PCB can't take the current for so much cores.
keep in mind that even not OCed ES do not made for lifetime.

 

[bayleyw]

1,5V over 6month was tested, nothing happened. i think its more the PCB can't take the current for so much cores.
keep in mind that even not OCed ES do not made for lifetime.

1.5V idle seems to be safe for the production parts but it seems the Rome ES parts have process differences (like how 2Sxx clock horribly compared to ZSxx), so the safe operating points might also be different. Did you test a Rome ES at 1.5V?