Overclock your AMD Epyc

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

I.nfraR.ed

New Member
Aug 20, 2019
19
31
3
Bulgaria
Hello!
I've updated the ZenStates app with Rome CPUID, however I haven't figured out the P-States (or manual overclock) yet.
There is no BKDG for family 17h and setting the well-known P-State MSRs doesn't change the CPU frequency. I have no access to NDA stuff.
If I set the P-State, then read it back it shows the newly set multiplier, but the actual frequency is still the old one. Same with the Linux script.
So I'm currently scratching my head over this.

It would help if you report that at least it detects the CPU correctly.
Performance Bias options "might" work, perhaps some of the others, but not P-State :(
I'm not even sure if the MSR are shared between Matisse and Rome.
Hopefully it won't "break" your system, but apparently I have no way to test.
EPYC1 shared the CPUID with desktop parts, but now Rome has a new CPUID.


Download zip from my github releases page: irusanov/ZenStates

Changelog:
  • Added Rome support (P-States won't work, but at least it should start without complaining about unsupported CPU)
  • Added support for 32-bit OS (tested on XP SP3 x86)
  • Reduced .NET Framework version to 4.0 (lowest I can go for the current codebase)

Update 21 Aug:
I have managed to change the frequency with commands to SMU. Unfortunately there's no public document describing all commands. I've updated the readme with what I've been able to find so far mostly by trial and error method. Should be able to provide at least manual overclock soon.

irusanov/ZenStates
 
Last edited:

MLyons

New Member
Aug 14, 2019
3
0
1
Thanks a ton for this. So in the bios there are some settings to change p-states that involve 2 values multiplied then divided by another. I can try to get a screen grab. I'll give the new version a go now.

EDIT: Ok so I've ran it and it grabs the CPU correctly. I set the performance bias to R15 and did an R20 ran and it seemed to make a bit of difference but I can't confirm due to it fluctuating at stock anyway. I tried to change the p0 state and I can't see much of a difference in score or clock speed being registered by hwinfo. I'm not sure if this is due to a power limit as I've been told we can increase how much power the CPU has up to 240W in a setting called cTDP but I'm yet to find the option. I'm trying is there anything else I can provide you that would be of help?
 
Last edited:

I.nfraR.ed

New Member
Aug 20, 2019
19
31
3
Bulgaria
What I've understood so far is that when in Overclock mode all limits should be lifted, except over-temperature protection.
I've released a new version with the manual overclock option and you can also control FID and VID.
I've currently locked DID down to 4 until I test the other options.

The thing is VID only works in offset voltage mode, so I'm afraid it won't do much on the server parts and that may limit the maximum possible FID (Frequency identifier). So, for example, if you engage manual overclock, all limits will be removed and you will be able to control the fixed frequency, however you will be limited by the initial VID (Voltage identifier), because it doesn't change when on Auto or Manual voltage override.
The CPU will probably auto-boost higher than what you will achieve in manual mode, because the VID will be the limiting factor.

If you can set a manual multiplier in bios, that essentially turns OC Mode on and higher the multiplier, higher the VID it boots with,
however I don't think there's such option on server boards.
Hopefully I can make VID work on full auto bios, not just in offset mode, but not sure if that's possible at all.

Thanks a ton for this. So in the bios there are some settings to change p-states that involve 2 values multiplied then divided by another. I can try to get a screen grab. I'll give the new version a go now.
If you're able to control the P-States, then maybe there's a chance if you set a high enough P0.
The frequency is determined by FID and DID, then there's also FID for the requested voltage.
FID is a frequency identifier from 0x10 to 0xFF.
The multiplier is then calculated with the following formula: m = (25 * fid) / (did * 12.5)
 
Last edited:

ari2asem

Active Member
Dec 26, 2018
745
128
43
The Netherlands, Groningen
maybe offtopic question

what is the difference between 1S en 2S engineering samples? not 1 socket or 2 socket, but this

2S1905A4VIHF4

1S1601A4VIHF4

@nero243 what motherboard are you using for your 1s sample? revision and bios version of mainboard?
 
Last edited:

epycforlife

New Member
Sep 21, 2019
3
0
1
Has anyone been able to manually overclock a Rome Epyc cpu?
I am looking to build a new rig, and would prefer to build a Rome based epyc build instead of threadrippper if I can increase the frequency a bit. Looking at the gigabyte mz32-ar0 as the mobo.
 

BryanH

New Member
Oct 28, 2019
9
0
1
I have a dual 7551 setup on a h11dsi-nt mb where the cpu water blocks restrict any airflow to the vrm heatsink. I believe this is causing the cpu to throttle down to about 1ghz under a long-term 96% load. I have already put a beefier heatsink on the mb chipset. I assume water-cooling the vrm would allow the cpu to run at a higher rate. Does anyone have any experience with a similar issue and are there any suggestions on a vrm water block? I haven't found a stock one to fit the 11.5cm footprint.

Thanks
 

alex_stief

Well-Known Member
May 31, 2016
884
312
63
39
What kind of load do you put on your 7551 for them to throttle this hard? I never experienced that so far, despite hitting them with "100%" load for several days. Have you checked VRM temps? Supermicro Super doctor can read them.
But I agree, if anything causes throttling with this CPU/mobo combination, it is the VRMs. In my setup, the VRMs run about 15° hotter than the CPUs themselves. With water blocks, you should have enough space left to aim a fan at the VRM heatsinks. Not a pretty solution, but it should work.
 

BryanH

New Member
Oct 28, 2019
9
0
1
What kind of load do you put on your 7551 for them to throttle this hard? I never experienced that so far, despite hitting them with "100%" load for several days. Have you checked VRM temps? Supermicro Super doctor can read them.
But I agree, if anything causes throttling with this CPU/mobo combination, it is the VRMs. In my setup, the VRMs run about 15° hotter than the CPUs themselves. With water blocks, you should have enough space left to aim a fan at the VRM heatsinks. Not a pretty solution, but it should work.
I run SAS (statistical analytics) and can control number of instances running concurrently and affinities for each instance. With 128 threads I would like to run 120 instances (15 per numa node) and it would be nice to keep them running around the 2.9 ghz turbo (I know that is wishful thinking). When installed the water blocks are about the same height as the heat sink essentially restricting it from any airflow. I have not used super doctor yet. I have downloaded it and tried to install but it keeps giving me grief so I shelved that program for now. My cpu's run in the high 30's under full load. When I overclock them to 3.3ghz they run in the high 50's. Regardless if they are factory set or overclocked the motherboard siren will go off and they will throttle down after a while. I have 3 other dual xeon computers and a very intricate water-cooling pump/filter/radiator set up for cooling which works great for the cpus but I don't think the motherboard can keep up thermally.
A66Z_132090421446955379Ec3hb2Di8h.jpg
 

BryanH

New Member
Oct 28, 2019
9
0
1
What kind of load do you put on your 7551 for them to throttle this hard? I never experienced that so far, despite hitting them with "100%" load for several days. Have you checked VRM temps? Supermicro Super doctor can read them.
But I agree, if anything causes throttling with this CPU/mobo combination, it is the VRMs. In my setup, the VRMs run about 15° hotter than the CPUs themselves. With water blocks, you should have enough space left to aim a fan at the VRM heatsinks. Not a pretty solution, but it should work.
test1.jpg
I got super doctor running and I ran 120 threads at 2.7ghz after about 90 seconds the vrm got up to almost 100c and the computer throttled down to 400mhz. This does not happen when I stress the cpu via cpuz.
 

Patriot

Moderator
Apr 18, 2011
1,470
802
113
Could you take a picture of your setup? I wonder if the waterblocks are putting pressure on the VRM sinks causing bad contact, in addition to not enough airflow. I also wouldn't expect cpu-z benchmark to max that many threads.
 

BryanH

New Member
Oct 28, 2019
9
0
1
setup.jpg sink.jpgvrm2.jpg
First off thank you for all the help, secondly sorry for the poor quality. I removed the sink from the motherboard to check for contact issues. As you can see from the thermal pad on the sink, there was full contact with the chips on the motherboard. The heatsink was not in contact with anything besides the 10 black chips between the cpu sockets. Are these 10 chips the only things that should be cooled?
 

alex_stief

Well-Known Member
May 31, 2016
884
312
63
39
You could at least test the "slap a fan onto the VRM heatsink" remedy. It is a time-tested method for this kind of problem: water-cooled CPUs with high power draw, thus very little airflow over the already flimsy VRM cooling solution.
Even if it is not a permanent solution for you, it is a quick way to check if better VRM cooling can help with your issue.

If I am not mistaken, the heatsink only makes contact with the chokes, not the mosfets. Which sounds really weird, so I will have to wait for an expert to shed some light on this.
 

BryanH

New Member
Oct 28, 2019
9
0
1
You could at least test the "slap a fan onto the VRM heatsink" remedy. It is a time-tested method for this kind of problem: water-cooled CPUs with high power draw, thus very little airflow over the already flimsy VRM cooling solution.
Even if it is not a permanent solution for you, it is a quick way to check if better VRM cooling can help with your issue.

If I am not mistaken, the heatsink only makes contact with the chokes, not the mosfets. Which sounds really weird, so I will have to wait for an expert to shed some light on this.
I am using a 2u environment which usually wouldn't allow for a fan on the heat sink but I was willing to move this rig to the top of the rack and custom vent the top with a case fan directly above the sink. However in a open case testing environment with a overpowered house fan blowing on the sink the problem still existed.
I don't know, I hope I'm not forced to live in 1ghz hell or have to only use 32 of the cores. This system is amazing with cinebench giving a score over 8300 when using 8 channel memory and clocking the cpu to 3.3ghz, it would be nice system if it could run 120 concurrent regressions at a decent speed for a extended period of time.
 

alex_stief

Well-Known Member
May 31, 2016
884
312
63
39
There is always running stock speeds. 2.4GHz is still better than 1GHz, but I know that's not what you are here for ;)
 
  • Like
Reactions: BryanH

BryanH

New Member
Oct 28, 2019
9
0
1
Anyone have the new username/pwd for Supermicro’s ftp site? I noticed a couple of days ago that the default was no longer working.
Only one I know of is "ADMIN" (all caps) for both username and password.

There is always running stock speeds. 2.4GHz is still better than 1GHz, but I know that's not what you are here for ;)
I agree. I removed the heat sink pad and used compound with a fan on the sink. I had to mill out a part of the sink because of a part of the mb that stuck up farther than the vrm (which could have been the problem in the first place). Was able to get a steady 90 degrees at 2.2ghz. I decided to get a vrm water block (Watercool HEATKILLER SW-X 80 DIY Ni (12003) ), ill let you know the results after its up and going.
Thanks for the help everyone.