i7-4790k or i7-5820k for FEA and Engineering Analysis

[cycleback1]

I need some help trying to put together another system beyond my current machine for primarily performing finite element optimization type tasks and other engineering analysis. I haven’t built a system for long time. Eventually the machine will become my day to day desktop (currently i7-860). I also need to buy a NAS or a storage server but I think that should be a separate box.

Primary Intended Usage:
==================
- Small 2D electromagnetic and thermal finite element simulations (generally < 20,000 elements) as part of an evolutionary optimization algorithm. For each generation of the optimization algorithm tens to a hundred independent designs are evaluated making the simulations extremely parallelizable favoring a large number of cores. I am not going to be running these all the time but currently when I do they generally take ½ day to 2 days to complete.

- Matlab simulations

- Solidworks: Relatively simple assemblies

- Rarely some larger scale 3D finite element simulations (generally < 250,000 elements). Maybe very rarely some CFD simulations.

-Eventually the machine will become my day to day desktop as my current machine is a i7-860 2.8 GHz, 8 Gb of RAM, ATI Radeon HD 5700 Series

-I don’t intend to over-clock the machine or play games on it.

The two systems I am thinking about are:
==============================
i7-4790k; $250
ASUS Z97-AR; $90
16 GB (2 x 8 GB) DDR3 1600; $160
Total: $500

i7-5820k; $300
ASUS X99-A; $235
16 GB (4 x 4 GB) DDR4 2400; $200
Video Card: Passive or semi-passive with Display Port and DVI outputs; $150
Total: $885

Components I have Already Purchased:
============================
Power Supply: Seasonic
Case: Fractal Design R5
CPU Cooler: Coolermaster Hyper 212 Evo
Storage: Crucial MX 100 512 GB SSD (Thinking about exchanging this)

Pluses for the i7-4790k:
Higher base (4 Gz) and turbo (4.4 GHz) CPU frequency than i7-5820k
Relatively in expensive through Microcenter (CPU: $250, MB: $90)
Built in graphics hardware


Pluses for the i7-5820k:
Two additional cores though they are slower
Twice the memory bandwidth
Maybe more future proof

Questions:
========
Which processor and platform is likely to be faster. The i7-4790k has quite a clock frequency advantage while the i7-5820k has two additional cores and twice the memory bandwidth. I know generally for large CFD type simulations a high memory bandwidth is highly recommended. I am not sure memory bandwidth matters much for small finite element simulations though.

Can the memory bandwidth of the i7-4790k and Z-97 be increased substantially by using a higher rated frequency memory and with reduced timings?

Is the additional cost of the i7-5820k and X-99 platform worth it?

I have been toying with the idea of trying to use an Intel Xeon Phi for parallelizing my small simulations further is it likely to work in a ASUS X99-A motherboard in the x16 slot?

What is the likely longevity of the Z-97 and X-99 platforms? I tend to keep computers for a very long time.

Is DDR3 likely to be phased out soon?

While DDR4 while theoretically more future proof I assume the timings and frequency are going to increase rapidly?

Is there a go to low cost video card that is passive or semi-passive with display port and DVI outputs?

Is the built in video card of the i7-4790k, HD 4600, okay for basic CAD work including Solidworks?

Any other thoughts or recommendations on what to purchase?

 

[Puffnstuff]

For what you are describing the socket 2011 based system would seem to be more appropriate.

 

[Ken g6]

Could the finite element simulations be offloaded to a video card with CUDA or OpenCL?

 

[cycleback1]

CUDA or OpenCL won't increase speeds in this case.

 

[Ken g6]

For what you are describing the socket 2011 based system would seem to be more appropriate.

Sans CUDA I agree. Another nice thing about LGA2011-3 is that, in a few years if the prices drop or your budget grows, you could upgrade to 10 or 14 cores.

I have been toying with the idea of trying to use an Intel Xeon Phi for parallelizing my small simulations further is it likely to work in a ASUS X99-A motherboard in the x16 slot?

A Xeon Phi should work in any PCIe x16 slot. Although if you can't use CUDA I'm suspicious that a Xeon Phi would work any better.

Is DDR3 likely to be phased out soon?

I hear DDR3 is supposed to be available at least through Skylake, which comes out next year. I don't know how much longer than that, though.

 

[Cerb]

https://forum.solidworks.com/thread/63854
Note that user's own results are filling the comments, including some nice Haswells on LGA1150.

If those can be considered representative of sims you might do, the cache and/or memory bandwidth seem to help as much as the cores, and the core scaling, while not amazing, looks to make up for lower clocks pretty well. Chances are, outside of such tasks, that take a substantial amount of time, everything else will be, "fast enough," if you have enough RAM. So, if those are representative, X99 all the way (basically: can they use up most of your logical cores on your current CPU? If yes, those benches likely relfect what you'll get).

Matlab is a question mark, but that performance will depend on what you're doing. Some things will be faster on the higher speed 1150 CPU, some faster on the wider 2011-3 one. Matlab also tends to be rather inefficient, if you can't spend time to optimize your work, but instead need to make it correct, run it, then throw it away or change it, so I really wouldn't worry too much about that (if you do have that time, and take it, you should also be able to speed up parallelizable parts, so no harm done with more cores).

 

[cycleback1]

https://forum.solidworks.com/thread/63854
Note that user's own results are filling the comments, including some nice Haswells on LGA1150.
.

Thanks for the link. This is the first quantitative data that I have seen.

 

[mfenn]

It seems to me that the first use case (parametric study using a (custom?) FEA code) requires the largest computational investment. Since your evolutionary algorithm relatively large generations of a few tens to hundreds of simulations, it should scale nicely with core count.

Overall, the 5820K provides about 25% more raw computational power than the 4790K. For a relatively long simulation time of 12-48 hours, that's 3-12 hours saved per run.

As for value, from a strict throughput per dollar perspective, the 4790K is better. However, if you place a premium on faster results, then the 5820K can pull ahead. Ultimately this is a value judgement that you have to make.