Milan VS Ice Lake
 

Hi All,

$10k budget, primary app is FloTHERM. Based on in-stock availability and pricing in my current location, I've narrowed my search down to one of these two options:

2x Epyc 7443 (24c Milan)

2x Xeon Gold 6366Y (24c Ice Lake)

Milan pros:
- ~15% higher ST performance (simulation initialization, radiation calcs)
- More cache

Ice Lake pros:
- Arguably better memory subsystem (see STREAM Triad benchmarks)
- AVX 512 (which I don't currently need)
- Dual socket Haswell system has had good scaling for us in FloTHERM. Dual socket EPYC 7301 (1st gen, 8 NUMA nodes) had poor performance...

Haven't seen a lot of discussion here on Ice Lake. Any recommendations?

Thanks,

Zachary

We do not have much data on Ice Lake Xeons in terms of CFD. 8380 seems to be about 30% faster than 8280 when used in OpenFOAM, however that is not a 1:1 comparison. I do not know how the FloTHERM load is on the CPU/Memory. Is it possible for the manufacturer to run some tests for you before you purchase?


If you do go with the Ice Lake option (btw I guess you mean 6336Y, not 6366Y?) then this article may be interesting to read:


https://www.phoronix.com/scan.php?pa...clear-21&num=1

If you had a really bad time with 1st gen Epyc CPUs with your particular software, the safe option would definitely be to stick with Intel for now.
Rome and Milan had some improvements to alleviate the worst-case scenarios for Naples. But they are still multi-chip designs, and you can't hide that entirely. Ice Lake is still a monolithic design, with all the advantages for software that is not quite ready for NUMA-heavy platforms.

What Intel did with the memory subsystem for Ice lake sounds really nice on paper and in stream, but I haven't seen it translate to real-world performance gains yet. IIRC, the Ice Lake results in the OpenFOAM benchmark thread were a bit disappointing. Don't get me wrong, they will get the job done. But I would have expected more based on the synthetic benchmarks. I guess huge caches still help AMDs CPUs a lot for CFD.

Unfortunately, the only way to be sure with this kind of niche software is to run some tests yourself. That would also factor in which models you typically run, and which parts of the workflow are most time-critical for you. If that's just not an option, stick with the safer solution - Intel.

Thanks, that is exactly what I wanted to know.

dual xeon 8375C would be a good option for 10K budget