Computer Specifications

Famous_Mortimer

What are the recommended specifications for a computer to run this program? Are you looking at supporting multi-cores in the future?

Famous_Mortimer

Curious what hardware others are running.
I7-4790K
16 Gb Ram
SSD hard drive
RTX 2080 graphics card

Victor

The internal matrix solver is already multi-core. I don't expect to make anything else use multiple cores.

The graphics card shouldn't have much effect on performance because the bottleneck for all the slow parts of the graphics is the CPU/RAM. Though people sometimes seem to bump into video memory limits - presumably with high resolution contour plots.

There can be an space advantage of a HDD over small SSD for big models that swap to disk since they can consume up to 100 GB or so.

Hengre

I am thinking to purchase a new computer for personal use.

Any suggestions from CAE/FEA community what to be considered for using CAD software an FEA with Mecway/Calculix, Paradiso, Intel MKL etc.
I mean are there some specific hardware demands to be noticed ?

Regards

prop_design

i would get the most memory you can and the fastest memory as well. i would also get the fastest m2 ssd you can. 8 cores is good. more than that doesn't seem to be of much benefit. but if you get 16 cores then you could still do other things while it solves. the larger the memory the larger model size you can solve.

anthony

Hengre

Thank you Prop_design

any ideas about graphic card?
AMD or Intel CPU?

prop_design

i'm an amd fan. however, your best to get a nvidia graphics card. they are supported more. for a cpu the amd zen3 would be the best. however, i'm not sure if you can actually get one yet. it might take a few more months. you'd have to check on that. i use a lowly laptop with an amd apu. i can solve things well. the run times aren't bad. my biggest problem is not enough memory and a hdd instead of a m2 ssd. so i have to keep my problem size way down. but run times aren't bad even on this old crappy amd apu. i've always been happy with the amd cpu/apu. but the zen3 is way better than the older designs. intel is just not doing well for awhile. so i'd stay away from an intel cpu. but that's just my opinion. i don't think it will matter much as far as which cpu or gpu. mecway doesn't need much for a gpu. but if you want to do rendering or solve fea/cfd on a gpu, then nvidia is the way to go for now.

prop_design

i should add a couple things that i have found looking at various vendors. they don't list the cas timing on memory. ddr4-3200 is good for amd cpus. however, some memory has lower cas timings which is what you want. so something like dell and hp usually use crappier memory. you also need to check the color gamut on the monitors. some monitors are really poor in that regard. dell and hp also use crappier ssd drives as well. even though they are m2 ssd they tend to use the ones with slower read and write speeds. i haven't found a vendor that really lists everything well. the custom builders are a little better about that, but still leave some important specs out. i'm not sure the best memory for intel cpus. i never look at those. but i believe it's a different frequency.

MikeMcMullen

I have been looking into this and have concluded that at the current time for MecWay and Calculix the video card makes little difference. Most of the prior development in to the types of solving libraries in the public domain or GPL are a little old compared to the current state of GPU design. This may not be the case for the high priced proprietary FEM packages. Earliest and most mature GPU compute libraries use CUDA, which is proprietary, though ROCm (open source) mostly from AMD may change this over the next few years. Increases in Pcie speed and GPU memory just now hitting the market may make gpu use more helpful, but consumer grade GPU's have poor double precision (DP) speed to avoid competition with the GPU vendors own compute products. For now you might get a slight improvement in behavior some times with big problems with any middle of the road Nvidia product. The AMD products may make this an interesting competition in a few years when the open source lowlevel compute libraries catch up.

I did a build for just this purpose earlier this year and used a Radeon VII. It is a beast at Milkyway at home with high DP speed compared to other consumer grade cards, but the software has trouble keeping it fed so it has significant idle time. Some folks in that community have luck running multiple problems simultaneously to cure this issue. Many FEM problems could be broken up to utilize current GPU's better but it is a low level programming task that is system and problem specific to optimize, so it will be a while. I agree with prop_design for the time being. Though Pcie 4.0 and a good capacity Pcie 4.0 nvme may help future proof your system. Note high end Intel systems with no more than 8 cores may be a bit better than AMD as their DP capabilities per core are somewhat better. MB's and processors that use more memory channels may be faster.