|
[Sponsors] |
![]() |
![]() |
#1 |
New Member
masofshad
Join Date: Feb 2022
Location: western USA
Posts: 4
Rep Power: 4 ![]() |
I'm trying to dig up more information about the memory timing options and behaviour for Epyc processors. Interest is primarily in developing an understanding to figure out which memory to buy. What has been your experience with memory timings?
Do Epyc or (and?) the motherboards for Epyc allow memory timings to be adjusted? Similarly, are the timings locked in some fashion by the CPU/motherboard or do they follow the JEDEC on the chips themselves? The behaviour that most people indicate is that the frequency of the memory is either the max for the CPU generation (2666/3200mhz) or the rating of the memory used; whichever is lower. Do the memory timings follow a similar behavior? Is there a min timings? -masofshad |
|
![]() |
![]() |
![]() |
![]() |
#2 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,426
Rep Power: 49 ![]() ![]() |
For the most part, primary memory timings can not be changed on Epyc platforms. You are stuck with whatever memory standard you bought.
Of course, I don't have experience with every single motherboard for Epyc CPUs. But the target market segment for these things usually means that such features are not exposed in bios. That's not limited to Epyc platforms, it holds true for pretty much everything in the server segment. It doesn't help that there are still no workstation-oriented motherboards for Epyc. That being said, it is sometimes possible to get these features back. I have a modified bios on a Supermicro H11DSI board that unlocks a lot of features which are hidden in the official bios versions. Primary and secondary timings being one of them. They can be changed below JEDEC standards, and the effect can be measured. As usual with overclocking, there is no pre-determined limit. Your results may vary. And to be honest, tuning towards the optimal performance on such a platform is extremely tedious and time-consuming. Doing it just for kicks is fine, but don't expect to hit ROI on that time investment ever. |
|
![]() |
![]() |
![]() |
![]() |
#3 |
New Member
masofshad
Join Date: Feb 2022
Location: western USA
Posts: 4
Rep Power: 4 ![]() |
Thank you for your quick response Flotus. I appreciate you sharing your experiences!
|
|
![]() |
![]() |
![]() |
![]() |
#4 | |
Senior Member
Join Date: May 2012
Posts: 551
Rep Power: 16 ![]() |
Quote:
Have you done any measurements that you can share? If I remember correctly there is an option for Threadripper in the Ryzen DRAM Calculator that may (or may not) be useful as a starting point for some moderately tighter timings on the EPYC platform? If you cannot control the memory voltage then I expect the possible range of improvement is rather minor. For kicks ofc not for ROI ![]() |
||
![]() |
![]() |
![]() |
![]() |
#5 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,426
Rep Power: 49 ![]() ![]() |
It's been too long, and I did not explore that route very thoroughly.
IIRC, I used one of the tamest profiles for 1st gen Threadripper CPUs from the Ryzen DRAM calculator. Of course, I already had to guess which type of memory available is closest to the Samsung RDIMM in my system. I measured performance with my own benchmark LB solver. Difference was in the low single-digit percent, but above noise. Optimizing further is not fun. Timings have to be entered in hexadecimal format, testing for stability takes ages with 512GB of RAM, errors can be hard to catch thanks to ECC, and the only recovery option with an unstable overclock is a full reset. So I left it at that. |
|
![]() |
![]() |
![]() |
![]() |
#6 | |
Senior Member
Join Date: May 2012
Posts: 551
Rep Power: 16 ![]() |
Quote:
Haha, well I can imagine it must have been a pain. I have my desktop in my cellar, in order to have minimal noise in my home office. It was a pain to test memory configurations and having to run down into the cellar each time it failed, in order to remove the motherboard battery. However, with only 16 GB of RAM, the startup is almost instant, so I took it as a good exercise of running the stairs instead ![]() Anyways, I am impressed that it worked at all. If it would have been possible to improve performance in the 10's of percent then it would most likely have been something that my workplace would consider. It is mostly time-consuming the first time and we have many workstations with identical components. Unfortunately it is gen 1 EPYCs. For our next round of new workstations I may have a look into custom BIOSes though ![]() |
||
![]() |
![]() |
![]() |
![]() |
#7 | |
New Member
The Frenchie
Join Date: Jul 2023
Posts: 1
Rep Power: 0 ![]() |
Quote:
Thanks in advance |
||
![]() |
![]() |
![]() |
![]() |
#8 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,426
Rep Power: 49 ![]() ![]() |
||
![]() |
![]() |
![]() |
Tags |
epyc, memory, timings |
|
|
![]() |
||||
Thread | Thread Starter | Forum | Replies | Last Post |
General recommendations for CFD hardware [WIP] | flotus1 | Hardware | 19 | June 23, 2024 18:02 |
Used Memory Accumulates During Course of Simulation Until interFoam gets Killed | Ship Designer | OpenFOAM Running, Solving & CFD | 7 | October 6, 2023 01:26 |
Ryzen 3X00 benchmarks and memory timings | Simbelmynė | Hardware | 11 | August 23, 2019 00:59 |
Lenovo C30 memory configuration and discussions with Lenovo | matthewe | Hardware | 3 | October 17, 2013 10:23 |