[Echidna]

I have an old workstation with 2x Xeon X5670 CPUs (3 memory channels each) and 96GBs (6x16GB) 1333 ECC RAM. I made a test to check if me memory is correctly populated and balanced. I ran an Ansys Mechanical benchmark with 6x16GBs RAM and next i removed one DIMM per CPU to evaluate the performance drop. I have to note here that the benchmark only consumes 32GBs of RAM so in both cases the memory was more than enough to avoid paging.

RESULTS (elapsed time)
6x16 = 676sec
4x16= 760sec

As you can see the difference is not that significant so i assume that my memory is not configured/populated correctly. There is a difference on the DIMM type though.
4 DIMMs are 2Rx4 PC3L-10600R-9-10-E2
2 DIMMs are 2Rx4 PC3L-10600R-9-12-E2

So the DIMMs are not exactly the same type but i am not sure this is the cause of the problem.
Anyone can shed a light here?

[flotus1]

I would disagree with the "not so significant" part. What you are seeing is a 17% (edit: now 12.5%) performance increase. I assume you used the maximum amount of cores available for this test.
As you mentioned, you ran a relatively small test. Generally speaking: the more memory a simulation consumes, the higher the benefit of more memory bandwidth.
An easy way to check if memory is populated correctly is by looking in the manual of your motherboard or the printing on the board itself. The DIMM slots are usually labeled A1, A2, B1 and so on.
Or you can boot up a windows version and use simple tools like CPU-Z and Aida64 memory benchmark to check if everything is ok. With CPU-Z you can also check the actual memory timings the system uses. But I guess it is the timings of the two slower modules.

Which motherboard do you have? Or what type of workstation?

[Echidna]

Originally Posted by flotus1

I would disagree with the "not so significant" part. What you are seeing is a 17% (edit: now 12.5%) performance increase. I assume you used the maximum amount of cores available for this test.
As you mentioned, you ran a relatively small test. Generally speaking: the more memory a simulation consumes, the higher the benefit of more memory bandwidth.
An easy way to check if memory is populated correctly is by looking in the manual of your motherboard or the printing on the board itself. The DIMM slots are usually labeled A1, A2, B1 and so on.
Or you can boot up a windows version and use simple tools like CPU-Z and Aida64 memory benchmark to check if everything is ok. With CPU-Z you can also check the actual memory timings the system uses. But I guess it is the timings of the two slower modules.

Which motherboard do you have? Or what type of workstation?

Flotus what is the actual difference of the DIMMs installed in my system?

[Echidna]

My motherboard is the Supermicro X8DTi-F latest bios update.

[flotus1]

Do you mean difference in terms of specifications?

4 DIMMs are 2Rx4 PC3L-10600R-9-10-E2
2 DIMMs are 2Rx4 PC3L-10600R-9-12-E2

The Timings are slightly different https://en.wikipedia.org/wiki/Memory_timings
With different DIMMs installed, The timings are usually set to the values of the "weakest link". But that should not have a dramatic effect on performance.
Which DIMMs did you remove for your test? The two slower ones or two of the four faster ones?

Edit: have a look in the manual of your motherboard, page 2-8 https://www.supermicro.com/manuals/m...0/MNL-1062.pdf
Here the optimal DIMM population for 6 and 12 DIMMs is described. As usual, populate the first slot of every channel with a memory module. So 1A, 2A and 3A for each CPU. The names of the slots should be printed on the board.

[Echidna]

I removed the 2 slower DIMMs

[flotus1]

That could explain a few percent in performance loss.
With 4 DIMMs the motherboard probably used slightly better timings than with 6 DIMMs.

[evcelica]

You want bad performance?
Try taking out 1 or 2 DIMMs on the same 1 CPU, which will unbalance the Memory. Or add that one you took out to the second rank of the first CPU.
The way you had it, with 2 in each CPU, is at least still balanced.

[Simbelmynë]

@Echidna

I don't fully understand. Why did you remove memory? It is a triple-channel memory and it is best to populate with 3(6) memory modules per CPU.

Check this link (especially the conclusion where he talks about CFD)

https://www.gamersnexus.net/guides/1...channel/Page-3