I’m thinking to get a machine for work specifically for FEA for PrePoMax, but I am not sure what to look for. Should I get a server with 128 cores at 2.5GHz with 3.5GHz max speed? or a desktop with only 24 cores at 3GHz max 6GHz. I’ve never maxed out the ram on our current machine so I am just focusing on the core count. I assume that faster ram would allow faster processor use so more significant core count? Unfortunately, we would need to run the prepomax test to know for sure how many cores would be ideal, but I have yet to see anyone saying that they are using more than say 20 cores efficiently. Right now I am using 16 cores.
And what about the RAM which is as important as the CPU; indeed if you’ve not enough memory, the system will swap = a drastic performance speed down
indeed, it’s recommended to use large memory as possible, after that is processor speed, number off processor also known could help to accelerate in Pardiso solver
I had a PC custom built for CAE in 2012, still working fine.
RAM wins hands down over n_CPUs & speed.
Main job in CAE solve is that you can fit jobs in RAM (in-core) 1st, then speed you don’t even care about if the model has been built correctly, you fine tune after. But running out-of-core (not enough RAM) is a total kill-joy.
2 Likes
OK, I see various people voting for more ram. But at least for PreProMax, I haven’t gotten to a point when I can reliably think to myself…hmm, this run will solve even if it takes 3 days. Like that situation borderlines in just go to the shop and use a force/strain-gauge to figure it out in an actual experiment. I exaggerate, but I usually run problems where I don’t want to spend more than 15 minutes waiting for an answer. specially if the answer is that the program failed to compute some where. I guess I should then I ask, how big/complex of a model do any of you run? what is the longest successful run you perform regularly?
As for the CPU speed, I think single-core performance is more important than multi-core performance if you are running a single job at the same time. If you run multiple jobs, this might be different.
1 Like
In FEA is very hard to summarise model size/complexity vs time.
Unfortunately, it is very application dependant, what kind of thing are you after?
Linear / non-linear, static / dynamic, contacts / no-contacts?
You can run a linear model (in general, dunno CCX) with 2M nodes in 20mins; or take a full day to solve a very non-linear one with a small mesh.
1 Like
CalculiX solver seems fast enough in convergences. I ran large number of bolt contact including plasticity in reliable times, it’s below one hours with only dual core processor and 16gb ram. Solver selection being used is Pardiso since PaStiX not stable at the day even known to be faster.
symetric load in tension like this took below half an hour to convergences.
2 Likes
I usually keep models as small and linear as possible. But now thinking ahead to new hardware, maybe ram would become a limitation on large assamblies. Our FEA machines I think has 128gb. It’s a z station model from 2015? I think. But the problem I see with thid hp model is that the 2ghz Zeon Gold processor which is relatively slow compared to our updated workstation model. Sure it’s got cores galore, but if I assign more than say 8 cores to calculix, the machine runs the problem significantly slower than the newer single CPU/32 core machines.
So that’s where my question is really coming from. If we buy machines with dual CPUs and end up with only ever using 8 or 32, then it would be a wasted investment. I would love to have a reliable benchmark like passmark where all CPUs have a nice ranking to compare… But based on PrepoMax.
probably it’s required testing personally about number of cores related to computational times at the same machine, from single core to maximum physical available. Different machine can lead to inconsistently due to memory type and processor speed. Solver selection between Pardiso and PaStiX also interesting to do compares.
btw, i used quadratic tetrahedral with default medium mesh size in previous example.
2 Likes
You know, with the new meshing features you could make most of these parts into transfinite hex meshes with higher accuracy and convergence rates, Improved computational efficiency, better element quality, reduced numerical diffusion, and their ability to handle large, structured models more effectively. That’s one of the drivers for me thinking about running larger models. From everyone’s points here, with larger models there is the inevitable longer solve time and larger ram needs. Secondly from what I understand, better single core performance for small models (quicker solve?) or larger core count for larger models (smaller local matrix chunks with larger more addressable global matrices ?). Also more core counts means you can run several independent parallel problems at once. That’s my take so far. Not super helpful but I guess it means that the passmark benchmarks may be a good indication of what CPU’s would be better for calculix/prepomax.
1 Like
With 128 cores, i would have a look to the “distributed memory” (MPI) capabilities of calculix (Pardiso, Pastix or other solver), compared to the shared memory ones (SMP i.e. multithreading with OMP_NUM_THREADS). I’m not a specialist, but i’ve ever heard that big models use less memory with a significant speed-up.
Note wer’e speaking here about the Calculix capabilities, not PrePoMax
Here is a thought I’ve done in the past to test solvers vs machines.
First, make sure that CCX doesn’t have a size limitation in itself for your cases.
Model a cube with tet10 elems, and keep increasing/solving higher mesh count on the machine you already have with say 4CPUs. You’ll hit a limit at some point not fitting in the RAM you have.
Take that largest model that fitted in your RAM now, and re-run it with 1CPU, 2, 3, 4 to 8 max.
Then plot times vs n_CPUs, you may be surprised. This may give you an idea if more investment makes sense, and/or CCX has some limitation anyways.
Unless an FEA solver has been tuned to use separate nodes (machines in the cloud), I’ve seen very few with a multi-CPU on same machine solve, that can speed things up enough over 8_CPUs or so. I’m gonna have a guess and say 4_CPUs will do, and above that on same model, I doubt you’ll get much speed up, and at some point it’ll even slow down.
I/O time (read/write) between the mesh ‘chunks’ split across n_CPUs plays a massive role, and CPU speed or RAM cannot do much about it at some point.
On the subject of CCX speed vs solvers, someone has done an in depth study, in case of interest:
When memory is large enough and number of processor too, Spooles MT also interested to test. However, it much depend on analysis case, I’m not sure it can be faster for large multipart contact.
Recently had to work on a client’s project and the model was not only large but had many contact pairs. We use Paradiso only and all our CPUs are Ryzens but different generations. Some of our Ryzens especially the AM4 ones didn’t even converge or diverge the simulation runs, but the AM5 pulled through. Then we found the problem, none of the AM4 Ryzens have AVX-512 support, only AVX-2. We had to scrap our 10+ years old Xeons which we previously used for Abaqus because they didn’t even have AVX-2 support.
So you need to consider CPU speed, no of cores, RAM, and also instruction set it supports. 7000 and 9000 series Ryzens are good choices currently. I didnt test latest Intel CPUs because they have two different core architectures and I am not sure how Paradiso will handle it.
5 Likes
Very interesting. I’ve though about buying up old servers to put together a DIY FEA machine at home.
there’s PaStiX solver or Spooles MT available and it seems old generation processor of AMD or Intel still usable even without latest AVX supports.
I consider PaStiX to be the best solver in PrePoMax. Is that not the case? Pardiso sometimes crashes.
indeed, it already know better in case highly nonlinear with large multipart contact and plasticity, thin and slender structure with nonlinear buckling also.
in previous version PaStiX has problem in convergences using damage material of MFront libraries, but Pardiso is not. This notice as not stable at the day, latest version known more stable but i’m not yet to test.