Dual EPYC 7601 on H11DSi-NT weak Cinebench R15 score

automobile

New Member
May 16, 2017
12
0
1
40
Hello, everyone! Can anyone help me in tuning of my H11DSi-NT to perform a decent Cinebench R15 score? Now I can only achieve a little over 3000cb...
Windows Server 2016
Windows power scheme: Performance
32 Gb of DDR4 2400 ECC RAM (Samsung) 2 modules x 16 Gb (one module on each CPU).
64 Cores, 128 Threads
Thanks in advance!
 

alex_stief

Active Member
May 31, 2016
672
213
43
35
I was also a little disappointed by the CB-score of my dual-7301: somewhere around 4100. But since I am neither using Windows nor cinebench-like Workloads I did not care too much.
The only reason I can see from the information you gave: not enough DIMMs. Each Epyc CPU needs at least 4 DIMMs (one for every die/memory controller) to get decent performance even in workloads that are not memory-bound.
What you should also check is CPU frequency under load. 7601 should run at up to 2.7GHz all-core for non-AVX code.
 
  • Like
Reactions: automobile

automobile

New Member
May 16, 2017
12
0
1
40
I was also a little disappointed by the CB-score of my dual-7301: somewhere around 4100. But since I am neither using Windows nor cinebench-like Workloads I did not care too much.
The only reason I can see from the information you gave: not enough DIMMs. Each Epyc CPU needs at least 4 DIMMs (one for every die/memory controller) to get decent performance even in workloads that are not memory-bound.
What you should also check is CPU frequency under load. 7601 should run at up to 2.7GHz all-core for non-AVX code.
Excellent! Thanks a lot! Added 6 memory modules (3 more to each CPU) and it jumped over 6000cb at once! Just G R E A T ! ! Thank you, man! You saved me - I checked every BIOS option, every Windows driver - no luck. I was even thinking that something is wrong with the mobo or CPU... Just addsome memory... Don't even know what to say! Lost for words ))
UPD: If anyone knows the best H11DSi-NT's BIOS/Windows performance settings for Dual EPYC 7601, please, share.
 
Last edited:

Krobar

Member
Aug 25, 2012
53
10
8
Excellent! Thanks a lot! Added 6 memory modules (3 more to each CPU) and it jumped over 6000cb at once! Just G R E A T ! ! Thank you, man! You saved me - I checked every BIOS option, every Windows driver - no luck. I was even thinking that something is wrong with the mobo or CPU... Just addsome memory... Don't even know what to say! Lost for words ))
UPD: If anyone knows the best H11DSi-NT's BIOS/Windows performance settings for Dual EPYC 7601, please, share.
Just another quick tip, you should benefit quite a bit more from 8dimms per CPU, those CPUs are dual channel memory controllers per die so 8 memory channels per CPU in total.
 
  • Like
Reactions: automobile

automobile

New Member
May 16, 2017
12
0
1
40
Yes, got all the 16 DIMMs on my H11DSi-NT populated and now it sometimes is jumping over 7000cb. Over 6500cb right away after the boot. It would have been even better if I had 2666 memory installed (I'm using 2400 16 Gb Samsung memory now).
Thanks to all for helping!
 
Last edited:

automobile

New Member
May 16, 2017
12
0
1
40
try turning off the hyperthreading in the BIOS, I believe they call it SMT or something like that. I get better performance with it turned off.
No. The Cinebench R15 performance dramatically decreases with "SMT" option (Simultaneous MultiThreading) disabled. It barely achieves 5500-5600cb (like a pair of my E5-2696V4). Are you sure you can improve performance disabling multithreading?! I beleive - more threads -> better performance. Although, I've heard that some CPU miners (for XMR, for example) are not optimized for multithreading and disabling that option could improve hashrate, but this is not our case. Cinebench R15 -> more threads -> more performance.
 

alex_stief

Active Member
May 31, 2016
672
213
43
35
SMT definitely helps with Cinebench. No idea where that came from.
Now if you want to do other stuff with your PC than running cinebench, disabling SMT can improve performance in some rare cases. But these are generally not the compute-bound scenarios.
 
  • Like
Reactions: automobile

EffrafaxOfWug

Radioactive Member
Feb 12, 2015
1,287
439
83
It's a very rare workload that performs worse under SMT these days, on either AMD or Intel hardware (or POWER, or any other SMT implementations).

When Intel first introduced SMT on their Pentium 4, there were some workloads that fared considerably worse because both execution threads assumed they should have exclusive cache access, and then spent a great deal of time thrashing the cache, resulting in poor performance. Since then there's been new processor instructions and software development to negate this sort of thing.

There was a recent article on SMT performance on an intel chip at phoronix; in every bench other than VP9 video encoding it was a net benefit (sometimes of the order of 30% improvement)
Intel Hyper Threading Performance With A Core i7 On Ubuntu 18.04 LTS - Phoronix

Zen is a somewhat more complicated beast by nature of its architecture (I can't find a decent on-stop shop review of it but IIRC from some reviews I read, some games and the windows scheduler have/had some problems with its SMT implementation) but overall the benefits are still largely positive for most workstation-esque workloads.
 
  • Like
Reactions: Tha_14

Evan

Well-Known Member
Jan 6, 2016
3,144
530
113
Power SMT is good because the choice of instructions is wide, ie in each core execution with SMT there is lots of possible executions, 16 pipelines within a single power8 core supporting SMT4.
Power9 SMT is even better.

Hard to really compare Power and Intel/AMD that are are doing 2 thread HT