AMD just announced the upcoming generation of Zen 3-based EPYC processors at CES 2021, and if you’ve built a computer recently, either for yourself or your business, you’ve probably had to answer the question — am I going Intel or AMD? For a number of years, Intel has been the clear answer, so much so that science and engineering software like Abaqus may partially rely on Intel-specific technology. You may not even find AMD in your vendor’s list of options to choose from. But recently, AMD has been making serious gains in their CPU business, capturing 8% of the x86 server market share, up from a miniscule 1% in 2017. Now it’s time for us engineers to start paying attention again.
AMD CEO Dr. Lisa Su shares generational improvements in their top-of-the-line Ryzen processor.
As a preview of what’s soon to come, I’ve run a few FEA benchmarks on my own latest PC build. This build is meant for content creation, 3D work, and high-end gaming, featuring the flagship CPU of the Zen 3 consumer line, the Ryzen 9 5950X. That’s a lot of cores — sixteen of them — and it puts this system in contention with the non-cluster Abaqus workstations we’ve been using at the office. We tend to think of the core count as the key number when it comes to how fast our FEA software solves, so why not find out just how well this latest AMD offering stacks up against the Intel performance we’ve taken for granted for years now? Maybe a few tests will give us an idea of whether AMD is ready to meet the Abaqus performance standard set by Intel.
Thirty-two threads of CPU power makes for a content creation and number crunching beast, especially in a highly overclockable configuration like you see here. But it’s important to note, however, that this self-assembled consumer-level machine does substantially differ from a preassembled consumer-level or enterprise-level machine:
- I was able to pick the best parts for my intentions…
- But I don’t get service and support from a system integrator.
- I can push the system beyond spec with CPU and memory overclocking…
- But I don’t get a robust warranty or potentially the same product lifetime.
- Consumer CPUs do not officially support error-correcting memory.
- Consumer CPUs have a far lower maximum memory capacity.
- Consumer CPUs have much lower memory bandwidth.
- Consumer SSDs survive fewer lifetime write operations.
Should you make a system yourself? Probably not. Should you get a high-end consumer desktop or a true workstation/server? That depends on what you’re doing and how much you’re willing to spend. But ultimately, I would not build a machine like this one just for the purposes of running Abaqus, simply because of how involved the process is.
I selected a variety of official Abaqus benchmarks and some of Caelynx’s own projects. I ran them on my Ryzen system and on some of the Intel systems I have access to at the moment. For the smaller models, I compared the Ryzen 5000 chip to a 10th gen Intel Core i7 in an 8-core vs. 8-core shootout. For the larger models, I compared the Ryzen 5000 to a Xeon Platinum Skylake/Cascade Lake chip in a 16-core vs. 18-core shootout. Ideally, I would compare the Intel Xeon line to AMD’s EPYC line, but the EPYC analogue to Ryzen 5000 isn’t quite out yet. From the twenty or so models benchmarked (which were chosen to represent typical Abaqus usage), I looked for any emerging patterns and found that, in all the Abaqus/Standard models I ran, there was indeed a performance trend.
None of these models took more than five hours to run, because at that point, it is my belief one should start looking into a different tier of machine (either GPU-accelerated or high core count). There is real value to short runtimes in terms of how fast you can make decisions and move projects forward.
|Model||Tire Inflation and Press||Welded Assembly Push Load||Battery Enclosure Natural Modes|
|Degrees of Freedom||729,264||3,187,841||703,599|
|Features||Simple wheel assembly, hyperelastic materials, large deformation, dynamic contact||Multi-part welded assembly, extensive dynamic contact, nonlinear materials including crushable foam||Large linear shell assembly with large solid masses|
|Scenario||Tire inflation & downward loading||Gravity load & push force||Natural frequency extraction|
|Model||Vehicle Seating Compression||Engine Block Bolt-Up||Gear Teeth Contact|
|Core Count||16 (AMD)/18 (Intel)||16 (AMD)/18 (Intel)||16 (AMD)/18 (Intel)|
|Degrees of Freedom||1,590,630||5,235,989||2,051,394|
|Features||Extreme elasticity, highly nonlinear contact, large displacement||Block casting components, nonlinear gasket deformation and contact||Multiple gears with fully modeled teeth, general contact|
|Scenario||Foam contact and compression||Assembly bolt clamping||Dynamic-implicit gear torque and contact|
Before We Get Into It…
An official Abaqus/Standard benchmark: Tire Inflation
There are many caveats to all the data here. I’ve done my best to control test conditions (fresh Windows accounts with minimal background processes, new/well-maintained machines, consistent Abaqus versions, multiple test passes), but these are not fully-controlled, like-for-like tests. There are many ways these systems diverge from component parity (memory capacity, memory channels, memory speed, storage speed, cooling solution, etc.). These differences are, in my opinion, very unlikely to reverse the trends presented here. More on this later.
It is also important to note that Abaqus solves different analyses with radically different methods. These solver methods use the hardware in differing ways, resulting in differing responses to changes in that hardware. Of all the Abaqus/Standard models I ran, when given sufficient memory, they all trended in the same direction. This does NOT mean that ALL Abaqus/Standard solvers perform better on the latest AMD than the latest Intel — but I didn’t find any evidence to the contrary. The bottom line is that these results cannot be freely extrapolated to solvers or problem scales that have been excluded from the comparison. In any case, the results presented here are consistent and they make sense, so I’m comfortable with them in that they confirm an ongoing trend in the CPU industry in general and, more importantly, show that that trend has fully taken hold specifically in Abaqus FEA compute.
With all that out of the way, what can the new Zen 3 processors do?
Prior generations of AMD’s Zen architecture have been known to solve more slowly their Intel analogues for Abaqus/Standard compute. This has been attributed to Abaqus/Standard’s usage of the Intel Math Kernel Library. On my systems, I have found this disadvantage to be somehow completely overcome. My consumer-grade, overclockable Zen 3 was highly competitive in raw core compute power, exceeding Intel in all the benchmarks I tried in core-to-core comparison. Regardless of the performance differential — which was often quite substantial — AMD CPUS are completely in the game when it comes to Abaqus/Standard.
As for the performance lead that AMD is showing over Intel, how could the system differences be impacting that? Indeed, part of this speed advantage for AMD may be attributable to the quality of the system I built around the CPU, namely the memory and the CPU cooler, but it is hard to imagine component parity reversing the Intel’s fortunes here. When it comes to memory, I (substantially) slowed down the Ryzen system’s memory to match the slower memory of the competing Intel systems, and found solve speed changed by only 10%. As for CPU cooling, only the 8-core Intel system had a potentially weak cooler, and it still reached within 5% of its max all-core frequency spec, which, furthermore, is not a continuous frequency state for an entire Abaqus run anyway. In a worst-case scenario, perhaps some analyses could push toward equal run times (like the welded assembly push load), but even that seems unlikely.
Dr. Lisa Su as Xeon and EPYC solve a weather simulation.
These less-than-scientific tests were very interesting to run and they showed some clear patterns, but I certainly would not make any numerical predictions based on them. We have to be very specific about what we can take away from these benchmarks. Here’s what I think:
- Regardless of Intel’s relative performance, AMD will likely be a great choice for Abaqus — with none of the old caveats. With Zen 3, their reputation for offering cheaper, more numerous, but notably weaker cores is a thing of the past. So is the notion that AMD CPUs should be restricted to Abaqus/Explicit usage.
- But AMD does look to have a lead against Intel in core-to-core processing power… for now. This lead may be exaggerated by the optimizations I’ve made on my AMD test system, but not by enough to reverse the hierarchy.
- The two companies are soon to be fighting hard for supremacy. Both of them have promising technology in the pipeline in the next couple years, and I expect Abaqus compute to make bigger leaps than we’ve seen in a while.
- The CPU isn’t the only thing that matters — you only see six benchmarks out of the twenty I ran because in many of them, the CPU was not controlling run time! (Some of them were also redundant Abaqus/Standard models that only confirm what’s shown here.)
- Consult your Abaqus VAR before making any purchasing decisions! These are complicated systems, and advice must be given on a personal level because everyone has different needs. That is something CATI/Caelynx certainly does for its simulation customers.
Talk FEA Software with the Experts
The CAE engineers at CATI/Caelynx would love to help you out with your software needs and concerns. We are practicing consulting engineers serving as official, certified support & training for the Dassault Systèmes products we use every day, such as Abaqus Unified FEA, CST Studio Suite, and more.
Reach out today to start a conversation.