workstation for hpc calculation for cae simulations

[Ing.Cecco]

Good morning to all,
I would like some advice on a possible workstation to buy for work for cfd simulations (star ccm+) and fem.
Currently we charge calculations to an external entity where we have rental clusters for a total of 32 core that for our work burden represents the minimum necessary.

Do you have any advice? number of cores? ram? graphics card? any standard models on the market? we will probably go on a solution with which the company has a consolidated relationship.

Thank you very much and good day.

 

[gil]

I'm not an expert. However, it has many proposals.
to evaluate the power you need and your buget. But 32 cores are many, xeon have 6 cores, I think it's hard to set up a workstation over the 2 processors so you'd stop at 12 cores or a little more.

 

[TAURUS77]

Good morning ing.
the latest intel xeon e5-2699 v4 (55m cache, 2.20 ghz min 3.6 ghz max) have 22 as core numbers 22 and 44 as number of threads so the problem of cores for distributed calculation is overcome as you can get to 44 physical cores and 88 logic.
the info that are missing in your post to better understand your needs are:
1) Since you have decided to buy the workstation you will use it in "local" and then connected with a small internal network (securely a faster solution with less bottlenecks) or not? ?
2) Are you going to use ws with a virtual machine for multiple applications? ?
3) the calculations you do for fem and cfd how many rams have occupied until now....in the new ws you can get to 1tb of ram...just add about 13000€ :smile:?
if I don't remember badly it takes about 1gb of ram each million degrees freedom as a general rule then depending on the type of solutor you use you can optimize the ram.
about a couple of years ago following a master's master of refinement a dandelion engineer explained how the constant need to do more accurate analysis in the cfd field carried with a certain frequency to change their ws and in particular the last one they had purchased was equipped with 512gb of ram.
4) for the video card I can only tell you that if the mesh of the analyses are quite complex and if you have to carry out control operations accurately for example how to enlarge a zone or a contour rotate and zoom in then it takes a waking vga....the card will be used with envy cubes or opencl to accelerate the calculation? Do complex or manipulated 3d modeling complex?
5) how do you intend to store and what do you prefer as a mass storage system for boot and os? ?

 

[Ing.Cecco]

Thank you. [MENTION=45643]bulus77[/MENTION] for the long answer. I actually stayed on the vague as I am not very experienced on hardware architectures for the hpc.
What I can tell you is that I am not actually bound to the workstation, but we could also work on servers as the machine would only do the calculations working locally connected to an internal network.

Currently we use the rental of computing power at an external body, in the following cluster configuration:

- 2 knots from 16 core (total 32 core)
- a knot consists of 2 cpu intel haswell 8 cores (e5-2630v3);
- 128 gb ram
- infiniband fdr interconnect

This configuration responds very well to our requirements even if having changed cfd software and no longer being bound by licenses that limit the use of cores, we could also evaluate a greater number of cores.

so we currently do all the pre-processing locally, we send to the external agency a file with which you can perform the calculations. post-processing is done in the enterprise cluster remotely or we can download it and work locally.

what I expect to do with the workstation or server that is, is something similar to what I do with the external hpc agency.

applications that would turn would be just cfd and fea (different software).

pushing on ram and graphics card wouldn't mind because I would also demand the mesh calculation and post-processing ratings.

Are these info enough?

I would need to know if buying an internal car is better than hiring hpc clusters.
the idea of expenditure I have in mind should not exceed 30 thousand euros.

Thank you very much.
Have a good day.

 

[TAURUS77]

Thank you. [MENTION=45643]bulus77[/MENTION] for the long answer. I actually stayed on the vague as I am not very experienced on hardware architectures for the hpc.
What I can tell you is that I am not actually bound to the workstation, but we could also work on servers as the machine would only do the calculations working locally connected to an internal network.

Currently we use the rental of computing power at an external body, in the following cluster configuration:

- 2 knots from 16 core (total 32 core)
- a knot consists of 2 cpu intel haswell 8 cores (e5-2630v3);
- 128 gb ram
- infiniband fdr interconnect

This configuration responds very well to our requirements even if having changed cfd software and no longer being bound by licenses that limit the use of cores, we could also evaluate a greater number of cores.

so we currently do all the pre-processing locally, we send to the external agency a file with which you can perform the calculations. post-processing is done in the enterprise cluster remotely or we can download it and work locally.

what I expect to do with the workstation or server that is, is something similar to what I do with the external hpc agency.

applications that would turn would be just cfd and fea (different software).

pushing on ram and graphics card wouldn't mind because I would also demand the mesh calculation and post-processing ratings.

Are these info enough?

I would need to know if buying an internal car is better than hiring hpc clusters.
the idea of expenditure I have in mind should not exceed 30 thousand euros.

Thank you very much.
Have a good day.

Hello!
I took a quick look at the configurator for a precision tower series 7910 with two intel xeon e5-2699 v4, 512gb of ram lrdimm a picture m5000 8gb and various storage options with boot pcie and raid 0 and various hd sas at 10000/15000 rpm and at 30000€ you will not arrive. .naturally excluded VAT and no discount that given the overall cost is obligatory.

 

[Andalo]

Maybe there are also supermicro systems. I refer to both the superservers (who come to 4 processors) and the superblades (where you go beyond, in a contained space).
that would be to be deepened. the only doubt is that the above systems do not meet the demand for the necessary graphics to mesh. that is an aspect that seems to me is supported by the tabs of the serious superworkstation, which also offer multiple pci-e 16x slots.
https://www.supermicro.com/products/nfo/xeon_mp.cfm:smile: