Finally: Overclocking EPYC Rome ES

[DaTruAndi]

So something is definitely weird about my overclocking attempts. e.g. I set the "All-Core Freq." to 3.50 GHz and in Core Temp and in CPU-Z it displays as the current Freq and Temp too. CPU utilization is 100% on all cores, but yet I get only half of or less of the "stock" performance.

Windows Task Manager shows a different frequency though. It starts at the set frequency but then jumps to another clock speed, typically maxing out at the published boost rate for this ES variant (2.9 GHz), but when I set higher clock speeds, it can even go down to 1.4 GHz.

I wonder if my BIOS is interferring and setting things differently?
It seems if I go above the supported boost frequencies, the system will get slower.

The good news is that the overclocking tool does "something" but the bad news is that it may not do what I want it to do.


If anybody here overclocked Epyc Milan hands on, would love to hear from you.

 

[mirrormax]

Have to check my BIOS settings, have not seen this yet.
As for the clock speed: I have not seen any tool touching the base-clock though.
RAM: Yes, 16 x 64GB

Don’t see a “power mode”. I only can do normal operation, or OC with some settings including changing the base clock speed up from 1600mhz. The other settings I am trying for figure out.
CCD Control : Auto, 2 CCDs, 3 CCDs, 4 CCDs, 6 CCDs
Core Control : Auto, ONE (1+0) , TWO (2+0), THREE (3+0), … SEVEN (7+0)

Also it seems stock according to bios the CPUs it runs at 1.1 V that may explain why the modification to run at 1.05 V resulted in lower total speed?

You need to dig deeper into the AMD cbs bios menus where you find those other settings.

 

[DaTruAndi]

You need to dig deeper into the AMD cbs bios menus where you find those other settings.

EDIT: I found it "Determism Control - Manual" enables "Determinism Slider - Auto/Power/Performance"
Some random links that help me to read up on a few things:

https://www.amd.com/system/files/documents/high-performance-computing-tuning-guide-amd-epyc7003-series-processors.pdf

https://lenovopress.lenovo.com/lp1267.pdf

Lenovo Documentation

Lenovo Documentation provides documents, publications and user guides for Lenovo server products and infrastructure solutions.

thinksystem.lenovofiles.com

But in the end, I have no been able to OC successfully. All monitoring tools "claim" that the clock speed/frequency is set to a higher number, but the benchmarks run slower.

I am not sure if ASRock Rack overrides all OC attempts or if the processor is just not OC able.
EPYC 7763 ES - 100-000000314-04

So hoping that someone who made it work for themselves can share the tool and settings they used.

 

[DaTruAndi]

Still messing with the performance of EPYC 7763 ES - 100-000000314-04, went through a lot of permutation of BIOS settings and settings in the OC Windows tool. Seen all sorts of supposed power draws and clock speeds, yet performance at best was CineBench performance bare metal Windows 11 was 87k-89k and typically when force stock max boost 2.9GHz for all cores and a voltage of 1.0-1.1
I also noticed that using 128 threads on the 128 core machine was very close in performance to using 256 threads. So hyperthreading really didn't add much for this type of benchmark.

Now interestingly, I tried ProxMox without mods/overclock on the OS level with one VM with 122*2=244 vcores (Cinebench does not start with 123*2 or more for some reason), and another VM with 63*2=126 vcores and let them both run simultaneously and got 55468 and 42656 and if the benchmark is linear, that would mean the combined performance was 98124.

 

[shpitz461]

I thought that there are differences in terms of the registers each cpu family uses, so an OC tool for Rome will not work (either at all or partial) on Milan. I think there was some tweaking done by the OC tool dev to make it work with Rome since it was originally developed for Naples.

 

[DaTruAndi]

I thought that there are differences in terms of the registers each cpu family uses, so an OC tool for Rome will not work (either at all or partial) on Milan. I think there was some tweaking done by the OC tool dev to make it work with Rome since it was originally developed for Naples.

I tried this GitHub - SKlayer/EPYC-Overclock
Which supposedly has a Milan release

also the OS and monitoring tools at least report a changed clock speed but if it is actually honored by the cores is a different question. Wonder if there is a tool that can monitor CPU throttling better?

 

[shpitz461]

Odd, the files say Rome but the release says Milan. Did you try to contact the dev?

 

[shpitz461]

Looks like the dev hasn't been around for quite some time: ' Last seen Nov 28, 2022 '

 

[mirrormax]

Still messing with the performance of EPYC 7763 ES - 100-000000314-04, went through a lot of permutation of BIOS settings and settings in the OC Windows tool. Seen all sorts of supposed power draws and clock speeds, yet performance at best was CineBench performance bare metal Windows 11 was 87k-89k and typically when force stock max boost 2.9GHz for all cores and a voltage of 1.0-1.1
I also noticed that using 128 threads on the 128 core machine was very close in performance to using 256 threads. So hyperthreading really didn't add much for this type of benchmark.

Now interestingly, I tried ProxMox without mods/overclock on the OS level with one VM with 122*2=244 vcores (Cinebench does not start with 123*2 or more for some reason), and another VM with 63*2=126 vcores and let them both run simultaneously and got 55468 and 42656 and if the benchmark is linear, that would mean the combined performance was 98124.

Yes it doesnt scale linearly up to 256threads how's your vrm temps?

 

[DaTruAndi]

Yes it doesnt scale linearly up to 256threads how's your vrm temps?

Measured how? I have a thermal imager, but the board has a heatsink on the VRMs. But no luck with any OC so far still

 

[mirrormax]

Measured how? I have a thermal imager, but the board has a heatsink on the VRMs. But no luck with any OC so far still

Have you looked at the ipmi sensors assume there's vrm temps there, at least there was on the supermicro and gigabyte boards I've used.

 

[DaTruAndi]

Have you looked at the ipmi sensors assume there's vrm temps there, at least there was on the supermicro and gigabyte boards I've used.

No I have not yet.

That said interesting that in Cinebench I had sudden improvements. Maybe because I found the "Power" setting in the BIOS?
I was also setting some of the other in the AMD CBS (package power limit and power draw related).
Either way, I went from around 75k Cinebench to about 85-87k. When I use the overclock tool and choose
All-core Freq: 2.9 GHz (which is the supportedly max boost of the ES chip I am having)
CPU Volt: 1.1V (I tried 1.0V and 1.05V too, not much difference, but going too low will throttle)
PPT Limit: 1500
TDC Limit: 700
EDC Limit: 700
Lock Freq: Checked
Dual: Checked

I get up to 93k-95k when lucky.

The ""Lock Freq." setting makes a real difference.

If I deviate from the settings, especially trying to increase the clock speed, the system gets slower at least in an all-core test.

Looking for another CPU bound test that updates the performance real time (as in FPS type) so I can play around with the parameters to see more immediate effects.
If anybody knows a multicore CPU test that keeps updating the performance - bet it hashes per second, FLOPS, or FPS, or whatever may be meaningful, I am all ears.

EDIT:
Now the "weird thing" is that some settings make the system "sound" and look as if it is working harder, e.g. the fans work overdrive and eg the "Core Temp" tool implies higher power draw, and the system works to try to stay cool. But the big put - even all the huffing and puffing the performance is less than at other settings.

 

[mirrormax]

No I have not yet.

That said interesting that in Cinebench I had sudden improvements. Maybe because I found the "Power" setting in the BIOS?
I was also setting some of the other in the AMD CBS (package power limit and power draw related).
Either way, I went from around 75k Cinebench to about 85-87k. When I use the overclock tool and choose
All-core Freq: 2.9 GHz (which is the supportedly max boost of the ES chip I am having)
CPU Volt: 1.1V (I tried 1.0V and 1.05V too, not much difference, but going too low will throttle)
PPT Limit: 1500
TDC Limit: 700
EDC Limit: 700
Lock Freq: Checked
Dual: Checked

I get up to 93k-95k when lucky.

The ""Lock Freq." setting makes a real difference.

If I deviate from the settings, especially trying to increase the clock speed, the system gets slower at least in an all-core test.

Looking for another CPU bound test that updates the performance real time (as in FPS type) so I can play around with the parameters to see more immediate effects.
If anybody knows a multicore CPU test that keeps updating the performance - bet it hashes per second, FLOPS, or FPS, or whatever may be meaningful, I am all ears.

EDIT:
Now the "weird thing" is that some settings make the system "sound" and look as if it is working harder, e.g. the fans work overdrive and eg the "Core Temp" tool implies higher power draw, and the system works to try to stay cool. But the big put - even all the huffing and puffing the performance is less than at other settings.

On Rome if I played around with settings I would often hit a point where performance went to 50% and I had to reboot to get back to 100%. Like it would just lock itself into low perf.

On another note I see there's epyc Genoa es chips on eBay now, anyone given those a go?

 

[DaTruAndi]

On Rome if I played around with settings I would often hit a point where performance went to 50% and I had to reboot to get back to 100%. Like it would just lock itself into low perf.

On another note I see there's epyc Genoa es chips on eBay now, anyone given those a go?

Would love a pair of 96 core monsters. Not sure what's the lowest is that the sellers would accept currently though. I tried with what I would be willing to pay but my offers were rejected

Other than that I have been scouring the web if anybody actually had positive results on Milan OC but have not seen anything really.
Should I have settled for cheaper previous gen Rome chips that is known to be overclockable rather than a Milan at the same core count? wondering.

 

[Spartus]

Well I have a pair of "Genoa 96 core monsters" and was coming here to check if this worked on them. No reports yet?

 

[mirrormax]

Well I have a pair of "Genoa 96 core monsters" and was coming here to check if this worked on them. No reports yet?

i would try with zenstates-rome for linux and see if you can look up the values first to see if it uses the same commands.

theres an --smu-test-message command to test sending smu commands

https://forums.servethehome.com/index.php?attachments/zenstates-rome-es-zip.14971/

 

[SHD_Tech]

Hey folks, I picked up some Epyc 7D12s on the cheap and PPT limit is hard locked to 105 with H11SSL-i 2.0. I have a buddy with an asus board that can hit 150w PPT without issue. From what I understand all I need to do is load the custom bios from executable fix and I should be able to push farther? These aren't retail cpus and some of the options in the epyc rome es overclocking tool work but a lot just say error on writing to smu.

Currently they boost to 2.3 all core (32c) but I am trying to get them any higher than that, ideally 3Ghz all core like my buddy is.

TDC/PPT bug does not work for me on stock bios it seems

Edit: Where do I find this bios? there was one in the other thread by nero, not sure if thats the one

 

[RageBone]

@SHD_Tech there is a link in the first post of this thread.
Not sure if they will help you though because the main modification was to get early ROME ES working and to get ROME support on the Rev 1.
AMD stacks multiple, usually 2 16MB large "Bios" images onto one single eeprom with the retail CPUs checking a few bits for compatibility.
The difference between Rev 1 and 2 H11XXX are 16MB vs 32MB eeproms.
With a Rev 1, you can there for only support one single CPU generation at a time by flashing its specific image.
Those early ES could not distinguish compatibility and would not post with code 00 or rather FF on SM.

If i remember correctly, AMD CBS etc options weren't patched in those images.
Maybe in the H11DSi ones.
But probably not.

 

[RolloZ170]

Hey folks, I picked up some Epyc 7D12s on the cheap and PPT limit is hard locked to 105 with H11SSL-i 2.0. I have a buddy with an asus board that can hit 150w PPT without issue.

check the AMD CBS power related settings of the ASUS board and diplicate it to the SM BIOS,
(RU.efi is a good tool for this without flashing BIOS i.e)
you also can unlock the AMD CBS menu in SM BIOS to make things easier.
( i can not help to unlock CBS menu, but there are guides for it, google is your friend)

 

[SHD_Tech]

@SHD_Tech there is a link in the first post of this thread.
Not sure if they will help you though because the main modification was to get early ROME ES working and to get ROME support on the Rev 1.
AMD stacks multiple, usually 2 16MB large "Bios" images onto one single eeprom with the retail CPUs checking a few bits for compatibility.
The difference between Rev 1 and 2 H11XXX are 16MB vs 32MB eeproms.
With a Rev 1, you can there for only support one single CPU generation at a time by flashing its specific image.
Those early ES could not distinguish compatibility and would not post with code 00 or rather FF on SM.

If i remember correctly, AMD CBS etc options weren't patched in those images.
Maybe in the H11DSi ones.
But probably not.

Ah yes you're right, under the spoilers. Got it and successfully flashed.

check the AMD CBS power related settings of the ASUS board and diplicate it to the SM BIOS,
(RU.efi is a good tool for this without flashing BIOS i.e)
you also can unlock the AMD CBS menu in SM BIOS to make things easier.
( i can not help to unlock CBS menu, but there are guides for it, google is your friend)

Looks like CBS is unlocked (kind of) and it does give me some additional controls, but I am still not able to push this cpu any harder than the stock rom with determinism set to power. Its nice having control of FCLK and UCLK controls now so I can see what makes a difference performance wise (which appears to be nothing lol). What I think I am running into now though is still a lack of power. Has me wondering what on this board is adhering to whatever that cpu is telling the bios that its power limit is. Shame voltage control isnt in there since that would give me some more leeway if it continues to refuse higher ppt/edc values.