|
[Sponsors] |
![]() |
![]() |
#1 |
New Member
Pierre LECHEVALIER
Join Date: Mar 2023
Posts: 29
Rep Power: 3 ![]() |
Hello,
I am about to buy a new CFD (Ansys Fluent in mono and multiphase simulation) calculation server for which we already decided to go with 2*AMD EPYC 9454 48C. The question is now what about the RAM, we need around 256 Go, our OEM offer us the possibility of : - 16*16 Go RDIMM @ 4800 MT/s single rank or - 8*32 Go RDIMM @ 4800 MT/s dual rank What would be the best option from your point of view ? Knowing that the EPYC 9454 has 12 memory channels, I would say let's go to fill as most as memory channels as possible, so go for the 16*16 option. But this single/dual rank makes me a bit confused... Thank your for your help ! Best regards, Pierre |
|
![]() |
![]() |
![]() |
![]() |
#2 |
Senior Member
Will Kernkamp
Join Date: Jun 2014
Posts: 365
Rep Power: 14 ![]() |
You are correct that you want all channels filled. That is very important because you want all those cores to work for you. Why not select 24x8 if cost is the issue. That is probably enough memory for your cases and it will be ~ 50% faster! If the OEM does not allow 24 dimms, you should select a different OEM.
There can be small difference in bandwidth based on single versus dual rank. Not sure how that works out on the 9545. If the OEM selects single rank for the 16 dimm config and dual rank for the 8 dimm config, maybe they know something about it. |
|
![]() |
![]() |
![]() |
![]() |
#3 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,426
Rep Power: 49 ![]() ![]() |
I hear you. I am currently involved in the acquisition of a few more compute nodes for our cluster.
And the OEM we are supposed to order from is no help at all. Populating memory correctly is the least of their problems. It seems to be tradition for OEMs to be confused for a few years, every time the amount of memory channels on CPUs changes. ![]() Anyway, for 2x Epyc Genoa, it's 24 DIMMs or go to a different supplier. If AMDs slides are anything to go by, they pretty much eliminated the performance difference for ranks per channel on DDR5. |
|
![]() |
![]() |
![]() |
![]() |
#4 |
New Member
Pierre LECHEVALIER
Join Date: Mar 2023
Posts: 29
Rep Power: 3 ![]() |
Dear Flotus & Wkernkamp,
First, thank you VERY MUCH for your reply, that's very clear ! Then, I think I do not have the choice, let's go to negociate with the OEM ![]() Best regards, Pierre |
|
![]() |
![]() |
![]() |
|
|
![]() |
||||
Thread | Thread Starter | Forum | Replies | Last Post |
EPYC 9004 memory bandwidth bottleneck threshold | Manuelo | Hardware | 5 | April 10, 2024 09:35 |
INTEL or AMD EPYC ROME for ANSYS Mechanical | evcelica | Hardware | 10 | March 25, 2021 15:27 |
Question on RAM | jhint9 | OpenFOAM Running, Solving & CFD | 2 | June 1, 2015 10:41 |
New workstation for different usage scenarios - CPU and RAM | natem | Hardware | 6 | August 7, 2013 02:47 |
memory (RAM) question | matt | FLUENT | 7 | November 28, 2006 03:30 |