[hurd]

Hello,

I am looking to get a new workstation for abaqus CAE and this forum seems to be a much more helpful place compared to our Abaqus support.

Here are the answers to the questions from the sticky post:Software: Abaqus Explicit with a bit of Abaqus Standard on occacion

1. License for 6 threads
2. Use case: quasi-static mechanical analysis with up to 1M elements
3. Budget: no fixed budget, don't want to waste money either
4. Users: 3 mechanical engineers in a medium company (~100 people, ~15 engineers)
5. Sourcing: We want warranty and support, so self-assembly is not an option.
6. Place: Germany
7. Extras: Would like info whether GPU would be helpful for my use case.

In fact I joined this forum when I read the quote from flotus1 "Abaqus standard solver is a weird beast. ... What's the "optimal" way you ask, and how to control it? I don't know, because neither does their support." because that is my feeling exactly.
The hardware stickie taught me about memory channels, which I had not considered before. I was thinking about buying a consumer CPU because of the few cores my license allows.

[wkernkamp]

Originally Posted by hurd

I was thinking about buying a consumer CPU because of the few cores my license allows.

I agree. It sounds like you already understand what to look for: 1) Memory speed and number of channels matter; 2) core frequency also helps. For a business environment, you probably want to look at DDR5 systems and the latest intel or AMD eight core systems. The two extra cores can manage interactions with the server from multiple users while the production calculations are ongoing.

[flotus1]

Extras: Would like info whether GPU would be helpful for my use case

As far as I know, GPU acceleration only works for Abaqus/standard. Explicit doesn't have it.
Unless you want to buy a used GPU, compatible models with noteworthy FP64 performance cost an arm and a leg. We are talking in the 5000€ and upwards range.
There is still the Tesla P100 16GB which can be had new for 1800€. Though if a system integrator still has that option available, and offers a similar price for it is questionable. And it's a server GPU with passive cooling. That's fine if you are actually looking for a server, but it won't work in a workstation-style case without a makeshift cooling solution.

For the CPU, I would go with the Intel Xeon W-3323. It's the latest gen with 12 cores and relatively high frequency. List price is around 1200€.
Why not latest-gen mainstream? You will want ECC support. And scaling on 6 threads will be better with this CPU.
Why not AMD Epyc? Abaqus/standard just plays nicer with monolithic CPUs. And for only 6 threads, you would not see too much of a performance difference even under optimal conditions.
Why 12 cores for only 6 threads? The HEDT/workstation CPU segment is in shambles. There just aren't any better options.

You probably know best how much memory you need for your models. The minimum viable amount for this CPU is 8x8GB DDR4-3200.

[hurd]

Thank you for your answers. Two more options that I would like to hear your thoughts on:

AMD 72F3 with 8 cores and 8 memory channels
AMD 5800X3D with 8 cores with 96MB L3 cache and 2 memory channels

[flotus1]

72F3: For explicit, it has been made obsolete by the similarly priced 7373X. Slightly lower core clock speed, but huge L3 cache similar to the Ryzen 7 5800X3D. For standard, both these Epyc CPUs are just about the worst-case scenario. They are made up of 8 chiplets, which means 4 NUMA nodes and high core-core latency. That's exactly the problem when trying to run Abaqus/standard.
The Xeon W-3323 strikes a better balance here. It might not be quite as fast for explicit, but being a monolithic design, it will be easy to run standard with consistently good performance. And it's significantly cheaper.
But sure, if you mostly care about explicit, and it doesn't matter how fast standard runs, AMD Epyc is the way to go.

5800X3D: it's a good budget choice. ECC support is technically there. But at 6 threads, it will already be limited by dual-channel memory. The Xeon will be significantly faster.
Think about how much you pay for the Abaqus licenses, and how much the engineers using the system cost. Saving 1500€ once by getting a slower computer is an objectively bad financial decision.

[hurd]

Thank you for bringing up the 7373X. That massive L3 cache might be enough to accomodate quite a few of our simpler models. That would offset any speed loss on Abaqus Standard because we do some much more in Explicit.

Now I have to get a quote and convince my boss . Will these processors get cheaper when the next generation arrives or will they just go out of stock?

[flotus1]

7373x_price.jpg
At least on the retail market, AMD server CPUs tend to get cheaper over time. Especially when a new generation launches. At launch, the 7373X was more expensive (over 4000$) than the 24-core 7473X.
But that's retail. The price you get from your system integrator is a different story.

[hurd]

Thank you once again. After some googling I was able to find two German suppliers that offer both the 7373X and the P100 in a complete system at retail-ish prices, so I am free to choose what I want .

[flotus1]

Two more things:
Buying a workstation from a German retailer, avoid this one: LS Computersysteme
Maybe they can do servers, but definitely not high-end workstations.

Abaqus/standard has hybrid parallelization, similar to MPI+OpenMP. Though their support won't tell you if it's actually that, or something that works in a similar fashion.
GPU acceleration works best when each GPU is associated with one "OpenMP" region of the solver. And "OpenMP" regions should not span across multiple NUMA nodes.
Lot's of jargon, the gist of it: If you want to do GPU acceleration, the Intel CPU would really make things easier. Or the other way round: GPU acceleration is probably not worth it with an Epyc CPU. You will be limited to a single NUMA node (i.e. 4 cores and 2 memory channels). And you will need to pester their support longer than I had the patience for, in order to find out how to bind the threads.

[hurd]

Thanks for the details about running with a GPU. With all those warnings, I think we are not getting our money's value when adding it. Combined with the cooling question, which both supplier contact people seemed to take a bit too easy, it might be wise to skip this option after all.