[SOLVED] Xeon Gold 5218R vs 6226R for HPC workstation

[warpino]

Dear all,

I am configuring a workstation for intensive 3D finite element numerical analyses in engineering.
Basically, it will consist of a dual socket motherboard with two Xeon Gold, either 5218R or 6226R, the latter for only 50 € more than the former.
Despite having less cores (16 vs 20 of the other cpu), the 6226R supports 2933 MHz as opposed to 2667 MHz of the other, and has quite a higher base
clock frequency (2.9 MHz vs 2.1 MHz).

Do you think 2 x 6226R cpus would be a somewhat better choice with respect to the 2 x 5218R solution?

Thanks for any suggestions

w

 

[kanewolf]

Dear all,

I am configuring a workstation for intensive 3D finite element numerical analyses in engineering.
Basically, it will consist of a dual socket motherboard with two Xeon Gold, either 5218R or 6226R, the latter for only 50 € more than the former.
Despite having less cores (16 vs 20 of the other cpu), the 6226R supports 2933 MHz as opposed to 2667 MHz of the other, and has quite a higher base
clock frequency (2.9 MHz vs 2.1 MHz).

Do you think 2 x 6226R cpus would be a somewhat better choice with respect to the 2 x 5218R solution?

Thanks for any suggestions

w

It is very difficult to write software that will scale to 32 threads (64 threads with hyperthreading). Have you benchmarked YOUR problem with that many threads to verify it scales?
The other thing to remember is that both of the CPUs you listed are 6 channel memory controller devices. That means that you want 12 DIMMs for maximum performance. So don't buy 4 32GB DIMMs, for example. You would get better performance with 12 16GB DIMMs.

 

[warpino]

It is very difficult to write software that will scale to 32 threads (64 threads with hyperthreading). Have you benchmarked YOUR problem with that many threads to verify it scales?
The other thing to remember is that both of the CPUs you listed are 6 channel memory controller devices. That means that you want 12 DIMMs for maximum performance. So don't buy 4 32GB DIMMs, for example. You would get better performance with 12 16GB DIMMs.

Dear @kanewolf ,
thanks very much for your reply.
I am not a developer, hence what I am saying here is mainly based on empirical evidence.
Actually, some of my problems do scale very well even on a (very) large number of cores. Some other problems, mainly to be solved using commercial software, do not scale that well and would certainly take advantage of a higher clock frequency.
Hence, I am looking for a versatile solution.
That being said, do you thing the 6226R might be a better option?

Best,

w

PS: I forgot to say... after some testing a dedicated GPU for HPC applications might be added to the system soon.
PPS: again about versatility, in some periods the machine might be devoted to run a couple of VMs for some commercial FEM applications we currently run on old Windows PCs

 

[kanewolf]

Dear @kanewolf ,
thanks very much for your reply.
I am not a developer, hence what I am saying here is mainly based on empirical evidence.
Actually, some of my problems do scale very well even on a (very) large number of cores. Some other problems, mainly to be solved using commercial software, do not scale that well and would certainly take advantage of a higher clock frequency.
Hence, I am looking for a versatile solution.
That being said, do you thing the 6226R might be a better option?

Best,

w

It would be my choice. You should also benchmark with hyperthreading disabled in the BIOS. You might get better performance with hyperthreadding OFF. But you will have to benchmark.
And 12 DIMMs to maximize the memory bandwidth to the CPUs.