I don’t see much talk about solve speed these days. I’ve been using the default Pastix solver and it seems to be just fine for the small things I analyze these days. However, I can only do part-wise analysis and it would be cool to be able to analyze larger assemblies or analyze using large deformations. Usually large deformations, contacts and plastic deformation analyses take up my computer for too long. To the point that I just kill it and give up. So I was thinking to upgrade:
I think I can get better hardware on my own, but I looked in Passmark CPU performance and this computer has a good performer (single thread CPU performance). I have another such computer, but its totally a dog, the single thread performance is terrible. It solves slower than my desktop.
Do you guys have hardware combinations that work great? I immediately went to the off the shelve because it has support. But maybe there is a vendor or a rig that makes much more sense to buy/make. Any tips/opinions would be appreciated. I was thinking that this would be a good place to ask since this is the only app that will be running on it. I might toss in Python in it too.
Wow. this was great! I ran the benchmarks on my computer and found that pastix will run better using 6CPUs instead of the 8 total in my desktop. So that really pulls me back to the question about how many CPU’s I need. I need around 8, everything else I can get is probably extra for running more than one analysis at a time. I had already heard of this and sort of confirmed it by self by observation but I had never ran a benchmark. Great link!
I found this other link and it has a really good rule of thumb:
Tip: A good rule of thumb when choosing hardware is to choose faster clock speed over quantity of cores!
Whatever computer I get, it has to have the fastest core possible. More RAM= bigger meshes, More CPUs= more experiments at a time.
its a little disappointing that one can’t just increase the solve speed by adding more hardware.
I’ve been buying AMD since early 2000’s. It was always a simple thing to go to fry’s and just grab a motherboard on sale. I don’t understand why fry’s died off. Anyway. I literally only want one thing, fastest performing single thread CPU and most memory supported. I go to intel’s site and instead of “hey! lets build a system, let’s start with the CPU!” I end up with “home? business? play?” WTF is that??? and sure, I could click on of those things and I did. I get to “oh we have awesome nucs for you!” but nothing to say, “this is the cpu I want”. I thought intel sold CPU! It looks like they sell multimedia now. “click on this brightly colored icon right here! here are another 20 brightly colored icons!” Whoever came up with that rat’s nest labyrinth is probably responsible for the current global downturn.
That said, does anyone have a direct link to a place to buy
Intel® Core™ i9-13900K Processor + 128GB ram + >1T SSD? or does PrePoMax support AMD processors fully? it runs just peachy with the AMD I run at home. I just don’t know if the newest ones are supported. This one for example has a similar single thread rating on passmark: AMD Ryzen Threadripper PRO 5995WX although very expensive.
Very interesting and useful guide, @FEAnalyst. It’s been a while since I want to perform such a benchmark calculation. I think I’ll post the translation of this post here in the forum so more people can use this amazing test. For my machine, more than 4 cores take more or less the same computational time. Unfortunately, I don’t have Pardiso solver installed (always find it very hard to do) to compare as well.
I would never buy such a thing for my own use. I don’t know who would. It’s definitely a few paychecks for me too. But my job has been pushing for more analysis and prepomax has been performing great. Like if something needs to fit in a tight space or if I need to have something flex a certain amount. Seriously Calculix has been such a great little tool. But things naturally turn more complex and I need to get results quickly. Sometimes I have to wait a while for results, only to comeback to failed or failed with results. So I want to get there quickly and restart quickly.
I have a HP Z820 with dual xeon and 128GB of ram which isn’t a new machine but can do plenty. It isn’t a machine you want to leave running if you have to pay for electricity
Much of the time the modeling/analysis approach that needs to be refined rather than just adding computing power (although more computing power and RAM can be helpful to a point). In earlier times we had to be more careful with model size etc so it is good to remember there are many things you can do to make faster running models on even rather moderate hardware.
I haven’t run a benchmark test, but the speeds seem to be quite good. The default solver seems to be very efficient at using multiple cores. Even when I only dedicate 24 processors, the resource monitor shows all 32 processors running close to 100 %. My system appears to be set up to overclock the processor at around 150%. I can also do other tasks, such as setting up another model, while an analysis is running without noticing a slow down for the other tasks.
In the past when I was using SolidWorks Simulation, I found that only one or two of the solvers used multiple cores, and they were very inefficient at using them. It seems that more cores may be more of an advantage with PrePoMax and Calculix.
We have a cluster we use to run CFD, and we have learned a couple of things:
Leaving some cores free to do traffic control and other tasks significantly increases speed over using all cores for the analysis. Based on the type of calculations required, this is probably even more important for FEA than for CFD.
Multiple “virtual cores” or “hyperthreading” does not improve performance for the types of calculations being done. Number of actual physical cores is much more important.
For CFD number of cores is more important than clock speed. I do not know if that is true for FEA.
I am not a computer expert, so this information may not be applicable in all situations, but hopefully there is something useful here.
You are absolutely right. From my point of view, what was stated on the Polish website completely disregards the difference between physical cores and virtual cores. Thus, 4 threads do not bring any improvement on a two-core CPU. No benchmark is needed for that knowledge!
For what I can say for CFD Analysis, memory bandwidth is very important. So, for example, if you are on a 64-core single CPU machine, the performance will be limited by the memory system and may be stuck at 16 to 32 cores. You are better off using two CPUs with 16 to 32 CPU cores each to double the cache and memory bandwidth. If this holds true for FEA, I don’t know, but it’s fair possible.
