[Cinek_Poland]

Hi. I will work in Abaqus and LS-Dyna using this software to compute stresses, strains and others physical quantity associated with mechanic , and metal forming. I will use 8-24 cores. For processor I can spend 1150$

I would like to ask , which processor should I choose ?
I thinked about Epycs: 7302p or 7313 also Intel Core i9 -10980xe .
I read that Abaqus and Ls-dyna using Intel MPI software and AVX 512 so will Intel processors have advantage ?

[flotus1]

The AMD CPU to buy would rather be the Epyc 7313p instead of the older 7302p. They cost about the same, 1000-1100€
If you are in for that AVX-512 support (and a monolithic CPU, which can come in handy for this software), my recommendation on Intel CPUs is the Xeon Silver 4316. It's the latest gen with 20 cores and 8 memory channels, albeit only at DDR4-2666. Cheapest listing I can find is 1060€. It's probably better than an I9-10980XE, at least for higher thread counts. I can't tell if there is a difference between having hardware support for AVX-512 for this software, or how high it is. Sorry

[AlexKaz]

If your problem size small, not more then 10000-100000 dofs, there is no advantages with latest Xeon 431x (I've bought two such cpus, with the serial numbers less than a hundred, i.e. these were the first processors that came out of the factory). I've tested LS-DYNA mpi with one Xeon 4314 with 8 channels up to 16 cores and Ryzen 5600x up to 6 cores with 2 channel. All CPUs have the same solution time on my problems.
Also i've tested problems with avx512. Solution time was more than with avx2, different is about 5-10%. Maybe it will be very nice to use avx512 with big proglem size, but I have not such problems.

Also I have second reason to use 5600x =) Xeon's cooler is very noisy. But Ryzen 5600x very cold with workable 5-6 cores at 4700-4800 MHz, and that's why his cooler very quiet.

[Cinek_Poland]

Thank you for your all answers.

I will talk with shop and I will ask if they will can test 2-3 processors in Abaqus benchmark. If I will result I will upload on this forum.

Mainly I have to solve metalforming process problems with or without heat exchange. If I'm using tools as rigid body and one deformable workpiece I need 100 000- 500 000 finite elements. But If Im using alsob deformable tool I need 2-5 milions finite elements and also many increments.

I found this https://www.amd.com/system/files/doc...s-explicit.pdf

I'm wondering why Epyc 72F3 is more expensive than 7313 despite of 72F3 has less core quaintity. (yes I know, epyc 72F3 has more cahe memory)

[flotus1]

I'm wondering why Epyc 72F3 is more expensive than 7313 despite of 72F3 has less core quaintity. (yes I know, epyc 72F3 has more cahe memory)

Mostly that. The 7xF3 Epyc CPUs have all 8 compute dies and their caches enabled, regardless of core count. Plus they have the highest core clock speed within the Epyc lineup.
You buy these CPUs when a 20% performance increase is worth thousands of dollars to you, because the increased productivity directly results in more revenue. Or if you spend lots of money on a software license that only allows you to run on a limited number of cores.

[Cinek_Poland]

Cache per core is very important in FEM processor performance ? Intel 3rd Scalable has less cache L3 per core than Epyc Milan and Rome.
Many people installing Intel MPI , after installation MPI software , simulations should be faster ?
does Abaqus favor Intel processors ?

[flotus1]

More cache is usually better, although there are a lot of caveats in that statement. But you can't compare cache size directly between entirely different CPU architectures.
Part of the reason why AMDs "Zen" CPUs have huge L3 caches is to offset the higher memory latency compared to Intel CPUs.

[Cinek_Poland]

Do you think Saphire Rapids should be more better than Epyc Milan ?

[flotus1]

Like I said already: I don't know.
Abaqus standard solver is a weird beast. There is usually more performance left on the table from running it sub-optimally, versus upgrading to the latest and greatest CPU. What's the "optimal" way you ask, and how to control it? I don't know, because neither does their support. And the fact that GPU acceleration works surprisingly well for this code has led me to believe that the CPU implementation is not as fast as it could be.
About LS-Dyna: never used it, so even less insight there