Wirkstation for CST and COMSOL Multiphysics

[Rox_2017]

Dear all,

I have to buy a workstation for the University where I work and I am looking for the best trade-off between performance and price offered by HP. Basically, I have to use on this computer the EM computational programs: CST Microwave Studio and COMSOL Multiphysics. According to the specific requirements of these software I have identified the following model:

Worstation:                  HP Z840

Processor:                    Intel Xeon E5-2643 v3

Operating system:    Windows 10 Pro for Workstation (4 Core Plus)

Memory:                       128 GB DDR4-2133 (8x16GB) 1CPU

GPU:                                NVIDIA Quadro P6000, 24 GB

Is there someone with experience in electromagnetic simulation with CST and/or COMSOL? If yes, I would be happy to receive suggestions about my choice. Could it be a suitable workstation?

Thank in advance for the answers,

Best regards.

[BambiBoomZ]

Rox_2017,

This is a tough one.  I hope other more informed than I will comment.

I don't have experience with either COMSOL nor CST,  but I've talked hardware a bit at my friendly neighborhood linear accelerator and within my second-hand limited knowledge of COMSOLand CST, there are a couple of parameters and quasi-equations that might act as l guidelines:

1. Memory: Simulation in general is very memory intensive and COMSOL projects apparently can quickly saturate the RAM, so both quantity and speed are important considerations. The technique to determine quantity is to run small models and monitor the memory used and the degrees of freedom from these smaller models. Perhaps the 128GB of RAM in your preliminary specification has been dervived from experience, but it may be wise to increase the quantity to prevent saturation.

2. CPU Cores: Again, my knowledge is limited,  but when I lent a system to a friend for some vonKarman turbulence simulations to be run on Matlab, we talked a awhile about the efficiency of parallelization in simulation software. With COMSOL, there is a relationship between the best effciciency of cores to the number of memory channels and the figure is 2 cores per memory channel.  If you have a dual Xeon E5 v3 or v4 dual processor, there will be typically 4X quad memory channels for a total of 16 RAM slots.  In the core /utilization realtionship, that means a 16X RAM slot motherboard is most efficient running COMSOL on eight cores.  Apprently addtional cores per memory channel can increase I/O latency.  There is an additonal latency introduced by the interlink of multiple CPU's.  I've seen this effect in visualization software. These benchmark have a relationship to simulation as rendering and simulation both, in effect, require a high calculation density in computing the positions and attributes of particles and polygons. I timed running fairly large sized- 3180 X 2140 renderings on an HP z420 with a single Xeon E5-1660 v2 (6-core @ 4.2GHz on all cores) and an HP z620 with 2X Xeon E5-2690 8-core  16-cores /32 threads)@ an all core clock speed of 3.4GHz.  The z620 is quirky as it has two memory channels for CPU 1 and only 1 channel for CPU 2. I always arrange dual processors systems to have symmetrical amounts of RAM between the two CPU's.The z420 was faster every time- and that equals three cores per quad memory channel instead of 5-1/3 cores for three channels-  plus bridging the two seperate CPU's. that bridging requires a one cycle parity check - hence the RAM is registered to maintin the two processor synch.  As one might imagine a one-cycle delay adds up over billions of cycles.

1. Consider the fastest possible single 8-core  Xeon:

8-core: Xeon E5-1680 v4 (8-core /16 thread @ 3.4 /4.0Ghz / 1.54TB RAM up to DDR4-2400:

https://ark.intel.com/products/92992/Intel-Xeon-Processor-E5-1680-v4-20M-Cache-3_40-GHz

This should be liquid cooled.  The Xeon E5-1680 v4  has the second highest average Passmark Single-Thread Mark of any Xeon E5 -2191 (1st is E5-1680 v2 at 2193)

This might be consideed as a 10 or 12-core or go to dual, higher clock speed 6 or 8-cores to increase the total calulation cycles per unit time but my recommendation is to first consider a single-CPU configuration.  However the lowered clock speed and core count to memory channel ratio may result in a degradation in a large project.

2. RAM: 128 or 256GB of DDR4-2400 ECC

3. GPU:  In both COMSOL and CST- simulation generally, a lot of video RAM is a benefit.  the difference toguh in the software is that COMSOL does not use GPU computing-e.g. NVIDIA Tesla  or Intel Xeon Phi, whereas CST  does use GPU acceleration.  If a Quadro P6000 24GB is in the budget, then yes that would be my choice.

4. Disk: 2X Samsung SM961 512GB M2. NVMe. drive 1 is the OS/Prgrams drive and Drive 2 is the Projects Drive. The storage drives: LSI 9300 MegaRAID SAS 9361-8i (LSI00417) PCI-Express 3.0 x8 Low Profile SATA / SAS  Eight-Port 12Gb/s RAID Controller running   2 to 5X HGST 4 to 8TB 7K6000 SAS In RAID 1 or RAID 50.

Something like that,...

BambiBoomZ

P.S. This is a very interesting project and if you know of a benchmark that would indicate performance in COMSOL and/or CST,  I would like to run it on my main system:

z620_2 > Xeon E5-1680 v2  (8-core @ 4.3GHz)(HP z420 liquid cooled) / 64 GB DDR3-1866 ECC / Quadro P2000 5GB / HP Z Turbo Drive 256GB M.2 (Samsung SM951) (OS/Programs) + HP 9212-4i RAID controller + Intel 730 480GB + 2X Seagate Constellation ES.3 1TB  The Single Thread Mark of this system is 2339.

> and compare the results to:

z620_1 > 2X Xeon E5-2690 (8-core @ 2.9 /3.8GHz)  / 64GB DDR3-1600 ECC registered / Quadro K2200 4GB / Samsung 850 Evo 250GB +Seagate Constellation ES.3 1TB  (The all-core clock speed is 3.5GHz)

> with the interest primarily in the compute core count to memory channel effciency.