der8auer overclocks EPYC 7601 with Supermicro H11SSL-i

Subsonic

New Member
Aug 10, 2017
12
4
3
34
I thought there is no way to overclock EPYC processors. Sadly he doesn't say anything how it´s done.

 
Last edited:
  • Like
Reactions: eva2000
Mar 28, 2018
32
3
8
He has been working on getting epyc to run on X399 motherboards but since it was a server motherboard in the video, it is most likely a motherboard vendor's development bios.
 

Patrick

Administrator
Staff member
Dec 21, 2010
11,971
4,932
113
If you have CPU and motherboard vendor support, just about anything is possible. With the ThunderX2 we were able to turn off the correct cores and clock everything correctly for basically any ThunderX2 SKU.
 

Kupaloid

New Member
Mar 16, 2018
18
5
3
53
Based on his newest video, he basically flashed i2c bus on the motherboard using special software.
 

Sapphiron

New Member
Mar 2, 2018
11
0
1
37
What is most interesting to me the massive performance drop going from 8 channel to 4 channel memory. I started a thread on thus forum asking if anyone had any data on this configuration. The consensus from most that the difference would be minimal. at least in Cinebench, that does not seem to be the case here.

Dropping from 5200 points to 3800 points. The CPU was clocked much higher than Epyc is typically clocked, but it is still interesting.

Also, not sure if his testing is really simulating a Threadripper 2990X. Are the CCXs on Ryzen Gen2 not significantly lower latency when using infinity fabric?
 

alex_stief

Active Member
May 31, 2016
666
211
43
35
He simply screwed up here.
The quad-channel configuration he tested was 2+0+2+0 memory channels per die, while the only sensible option for Epyc would be 1+1+1+1.
Many people -including me- tried to make him aware of this mistake, so he tested again...with the same suboptimal memory configuration in various flavors :rolleyes:

What we can still take away form this: it is entirely possible that TR 2990x only allows this kind of memory configuration where two dies are not directly attached to memory. The performance impact would be catastrophic.
 

alex_stief

Active Member
May 31, 2016
666
211
43
35
Dropping 25% performance in Cinebench, a benchmark that has become popular for high core count CPUs because it is not very sensitive to memory speeds. I stick with catastrophic.
Cinebench is more of a best-case scenario here, other workloads will suffer even more.
 

Patriot

Moderator
Apr 18, 2011
1,311
695
113
Dropping 25% performance in Cinebench, a benchmark that has become popular for high core count CPUs because it is not very sensitive to memory speeds. I stick with catastrophic.
Cinebench is more of a best-case scenario here, other workloads will suffer even more.
Cinebench is INCREDIBLY sensitive to memory clock and latency.
 

alex_stief

Active Member
May 31, 2016
666
211
43
35
Is it really? All benchmarks I have seen so far indicate that it scales very poorly with memory speeds. It scales a little better on Zen CPUs for other reasons, but still far from being memory bound. Sure, If you want those 50 points to get first place in an overclocking competition you have to crank up the memory as well. CPU clock speeds translate almost 1:1 into higher scores, memory speeds do not.
 

Sapphiron

New Member
Mar 2, 2018
11
0
1
37
@Sapphiron - Just tweaking the piece now EPYC 1, 2, 4, 8 channel memory. Odds are it will be published no later than Thursday but as soon as tomorrow.

As a note, we are using quad channel memory one DIMM per die, not quad channel two DIMMs per every other die.
Looking forward to it.
 

Subsonic

New Member
Aug 10, 2017
12
4
3
34
Using the 3M Novec Fluid for cooling is visually such a cool thing. Look at minute 11 for some serious nerdporn :-D