Every couple years there is a hot new technology that gets a lot of press, and clients are interested in learning more about how this technology may help them do more engineering.

Recently, there has been some interest in the latest AMD EPYC “Rome” processors with up to 64-cores per processor.

At the time of this writing the latest Intel technology is Cascade Lake Xeon Refresh which brings an incremental improvement to Intel Cascade Lake, at pricing in line with AMD. We tested the latest Intel/AMD available on Cloud, as well as on-premise HPC using a couple popular CAE codes (ANSYS LS-DYNA and Dassault Systemes Abaqus) to evaluate the overall performance at the application level.

ANSYS LS-DYNA

ANSYS LS-DYNA  is a general-purpose finite element program capable of simulating complex real world problems. It is used heavily for crash in the auto and aerospace industries. TotalCAE ran the popular car2car benchmark on EPYC and Intel based systems to see the performance difference.

car2car Benchmark Model

The car2car benchmark from topcrunch uses a NCAC minivan model created by Dr. Makino and analyses a head-on collision between 2 vehicles with around 2,500,000 nodes and elements. This model was chosen as it has been widely benchmarked, even though it is smaller by today’s standards.

The AMD system tested was on Azure (HB120rs_v2) with AMD EPYC 7V12 64-Core Processors at 2.45 GHz. Since Intel Refresh is not available on cloud yet, we used Intel Xeon Platinum 8168 24-core core processors at 2.7 GHz on Azure ( HC44rs) and used the list cloud pricing.

The results of the two benchmarks are below.

car2car on Intel Xeon (Skylake)
car2car on AMD EPYC (Rome)

You can see that 120 cores of AMD EPYC takes a bit longer than 88 cores of Intel Skylake. So even though they have similar clock rates, the Intel based solution takes less time on less cores.

Since most CAE software (like LS-DYNA) is licensed by core, being able to use less cores results in a savings on the software licensing. If you include the approximate on-demand costs of the software with on-demand compute pricing, running on more cores of Intel based HC44rs ends up with a lower overall job cost than AMD EYPC HBv22

Approximate on-demand job cost with licensing and compute Azure HC44rs (Intel)
Approximate on-demand job cost with licensing and compute Azure HBv2 (AMD EYPC)

Dassault Systemes Abaqus

For this benchmark, a smaller real-world Abaqus model was utilized and ran on an on-premise HPC that has AMD 7742 processors (64 core @ 2.25GHZ) with 512GB of memory.

The same model was run on-premise with dual Intel 6226R ( 16 core @ 2.9GHz Intel Refresh) with 192GB of memory.

Similar to the LS-DYNA results, The Intel system took less cores to achieve a lower runtime for Abaqus from the graph below (24-core of Intel vs 60 core of AMD below)

If you include the licensing, the Intel 24-core job required only 19 Abaqus tokens to achieve the speedup, while the AMD EPYC at 60-core required 28 tokens. A savings of 9 tokens per job is a significant cost saving for the Abaqus licensing.

It’s Not Just About the Benchmarks

TotalCAE is responsible for everything from when the user submits their simulation job to be solved, until they get their results. For this reason TotalCAE favors adopting stable and proven technologies in our HPC managed client clusters, and on-demand cloud solutions.

If the users job crashes, has numerical errors, or there are other problems related to new technology, it doesn’t matter how good, fast or cheap benchmarking looked. Some criteria TotalCAE uses when evaluating new technology include:

  1. Has the solution been in the marketplace a while and had the kinks worked out by others. We don’t like to be the guinea pigs for new technology for production systems that prevent engineers from getting work done.
  2. Is the technology in widespread use by the CAE software vendors, so we can work with ISV on resolving model issues, to keep engineers working.
  3. Is the solution well supported in the field by the Linux (Red Hat) operating system, to avoid discovering stability problems that interrupt jobs.
  4. Does the solution perform much better than existing solutions at the application level to save our clients time and money.

Based on our benchmarking of price/performance and criteria for new HPC technology, we still are favoring Intel Xeon based HPC systems for commercial HPC production systems. TotalCAE will continue to keep an eye out on the space.