$5000 Desktop for Simulations (potential GPU-Acceleration)

M

MirageMobile

Limp Gawd
Joined
Jan 22, 2007
Messages
489
I am looking to build a machine to assist with my graduate education. Any help would be greatly appreciated.

1) What will you be doing with this PC? Gaming? Photoshop? Web browsing? etc

The sole application is large scale simulations (for a local machine).
Problems to solve will involve:
computational fluid dynamics
multi-physics simulations
computer aided design
matlab scripting

2) What's your budget? Are tax and shipping included?

$2000 - $5000 is my initial guess of the total system cost including tax and shipping

3) Which country do you live in? If the U.S, please tell us the state and city if possible.

USA

4) What exact parts do you need for that budget? CPU, RAM, case, etc. The word "Everything" is not a valid answer. Please list out all the parts you'll need.

CPU (Interested in dual CPU, but have zero experience here. Need advice)
MOBO
RAM
GPU (Telsa K40 consideration)
PSU
SSD (boot) and HDD (storage of~20 TB)
CASE

6) Will you be overclocking?

No

8) When do you plan on building/buying the PC?

Now

9) What features do you need in a motherboard? RAID? Firewire? Crossfire or SLI support? USB 3.0? SATA 6Gb/s? eSATA? Onboard video (as a backup or main GPU)? UEFI? etc.

DUAL-CPU
RAID (I have zero experience but may need this. Need advice here)
USB 3.0
SATA 6GB/s
 
You're way underestimating your final price.

Choose dual processor, or high-end Nvidia GPU. You can't have both for $5000. The GPU alone is $3000!

I assume you're targeting Nvidia Kepler because you need the double-precision FP?

Also, how important is ECC memory to you?
 
What programs will be used?
Check what the programs can use and what is recommended by the programs.
It could be that the programs don't need double-precision and that a K40 is just totally useless. It could be that they don't use more than 8 cores, so a dual CPU wouldn't be beneficial.
 
He could also be running custom simulations for graduate work (hinted by Matlab). So no system requirements available :D

That's why I asked about his FP64 needs before committing. Also, OP: you do know that to get a performance improvement from GPGPU, you need to have large quantities of independent data sets to be operated on, right? Just like the pixels of a 3D scene, each of which can be rendered independently. So if you have dependencies in your data sets, you might be better off with more cores.

It would be nice if the OP actually replied. But he's a once-a-week poster, so we'll have to be patient.
 
The first task I want to perform is image post-processing by particle image velocimetry (PIV) analysis, where I generate velocity vectors of moving particles in a fluid via interpolation between consecutive images (typically 30,000 images per set and I will generate over 100 sets). See the two images at the bottom of this post for a visualization. My goal is to to implement GPU acceleration into the in-house post processing matlab scripts, which I have not done before.

FP64 Needs

Unfortunetly, I have not performed this type of analysis before. However, literature suggests that significant time savings can be achieved using GPU-accelerated image post processing.
My goal (because the money is available today) was to build a system capable of GPU acceleration now.

The delay in my response is because I am currently working on defining if I need double-precision. I will repost back when I have confirmed.

Below are references indicating good performance increasing with the GPU acceleration

1. http://ltces.dem.ist.utl.pt/lxlaser/lxlaser2010/upload/1670_lfplgn_3.4.5.Full_1670.pdf
2. Cambridge many-core.group - Tomographic PIV


Therefore, lets proceed with the assumption of a $5000 budget, single-precision is sufficient, and GPU acceleration is desirable.


ECC Memory

Correct me if I'm wrong, but ECC memory appears to be of an equal price (looking at 16gb x 4 kits on newegg) and appears to be of equal performance (http://www.techspot.com/article/845-ddr3-ram-vs-ecc-memory).
Therefore, the ECC requirement would drive the CPU/MOBO selection into Xeons?
If so, then ECC sounds fine to me.


Other Programs

I would like to simultaneously run:
1. Other in-house Matlab programs (I would need to implement GPU acceleration into these matlab scripts)
2. Commercial CFD software (Ansys, Star CCM+, NEK-5000) (these programs contain GPU acceleration options)
3. Solidworks 3D CAD

Example results

Time%20Resolved%20PIV%20(TRPIV)%204.jpg


PIV_through_stagnation_flame.jpg
 
Yeah, image processing is exactly the strength of the GPU (mass quantities of mostly independent data). So you just need to decide between 32-bit or 64-bit accuracy, as we're talking the difference between $3000, or $1200 for a Titan X.

I would recommend the Titan X over the cheaper GTX 980 Ti only because it has 12GB memory, and memory size is often the limiting factor in getting simulations running.

The only thing the Titan X doesn't have compared to it's $4000 Quadro M6000 counterpart is ECC RAM.

Also, I found this article interesting: The usefulness of ECC memory on GPUs is of limited value, since errors don't tend to occur. Might save you some cash.

Is the ECC Performance Price Worth it for GPUs?

Some interesting things to consider there!

Until the GTX 1080 releases next month, Titan X is the fastest 32-bit Nvidia GPU you can buy. But the 1080 will only have 8GB ram, and will probably be in short supply, so not very good.
 
Last edited:
Assuming 32-bit no ECC GPU is fine, you have more than enough to get a dual processor system. The 12-core x 2 Broadwell-E is a pretty good value.

PCPartPicker part list / Price breakdown by merchant

CPU: Intel Xeon E5-2650 V4 2.2GHz 12-Core Processor ($1115.99 @ SuperBiiz)
CPU: Intel Xeon E5-2650 V4 2.2GHz 12-Core Processor ($1115.99 @ SuperBiiz)
CPU Cooler: Scythe Kotetsu 79.0 CFM CPU Cooler ($39.89 @ OutletPC)
CPU Cooler: Scythe Kotetsu 79.0 CFM CPU Cooler ($39.89 @ OutletPC)
Motherboard: Supermicro MBD-X10DRI-O EATX Dual-CPU LGA2011-3 Motherboard ($415.98 @ Newegg)
Memory: Kingston 64GB (4 x 16GB) Registered DDR4-2133 Memory ($336.99 @ SuperBiiz)
Memory: Kingston 64GB (4 x 16GB) Registered DDR4-2133 Memory ($336.99 @ SuperBiiz)
Storage: Samsung 850 Pro Series 512GB 2.5" Solid State Drive ($216.63 @ Amazon)
Storage: Western Digital BLACK SERIES 4TB 3.5" 7200RPM Internal Hard Drive ($196.99 @ Amazon)
Video Card: EVGA GeForce GTX Titan X 12GB Superclocked Video Card ($1023.99 @ B&H)
Case: Antec P380 ATX Full Tower Case ($158.99 @ SuperBiiz)
Power Supply: EVGA SuperNOVA P2 1200W 80+ Platinum Certified Fully-Modular ATX Power Supply ($187.00 @ Newegg)
Total: $5185.32
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2016-05-11 00:18 EDT-0400

If you need more than 24 cores, you can upgrade this system to 2 x 22 = 44 total cores. I left enough PSU space to upgrade to 150w TDP CPUs AND add a second Titan X.

But I chose the Kotetsu because it has a very slim and tall profile to make room for those stacks of memory, but still can handle > 100w CPUs quietly.

If you have to get the 64-bit GPU, you'll have to cut this down to a single-socket.
 
Last edited:
980Ti is frequently faster than the Titan X for acceleration and it's very hard to find a program that chews up more than 6GB of video Ram.
Out of all GPU's currently available, (tested in video processing programs), the 980Ti is the best value.
I would opt for a case without a window and better noise isolation.
But appart from that I support the build above.
 
Defaultluser,

Thank you for the taking the time to assist me! I am still working on the FP64 requirement.

I do have some questions,

1. I need ~20 TB of HDD storage included in the $5000 budget. Selecting RAID5 with 4 x 8 TB yields ~24TB. This will cost $1280 with the Western Digital Red 8TB 3.5" 5400RPM Internal Hard Drive.
- To make room in the budget, would you suggest removing the 2 x E5-2650 v4 and replacing it with 2 x E5-2630 v4 ($630 each)? Intel® Xeon® Processor E5-2630 v4 (25M Cache, 2.20 GHz) Specifications

2. Please provide insight into the suggestion to use 1 x Titan X vs. 2 x 980 TI? The SLI setup has more cores but 24/7 long term reliability is more questionable and power reqs are higher? Or is this a matter of avoiding SLI?
 
MirageMobile said:
Defaultluser,

Thank you for the taking the time to assist me! I am still working on the FP64 requirement.

I do have some questions,

1. I need ~20 TB of HDD storage included in the $5000 budget. Selecting RAID5 with 4 x 8 TB yields ~24TB. This will cost $1280 with the Western Digital Red 8TB 3.5" 5400RPM Internal Hard Drive.
- To make room in the budget, would you suggest removing the 2 x E5-2650 v4 and replacing it with 2 x E5-2630 v4 ($630 each)? Intel® Xeon® Processor E5-2630 v4 (25M Cache, 2.20 GHz) Specifications

2. Please provide insight into the suggestion to use 1 x Titan X vs. 2 x 980 TI? The SLI setup has more cores but 24/7 long term reliability is more questionable and power reqs are higher? Or is this a matter of avoiding SLI?
Click to expand...

That's a good alternative, I forgot how much cheaper the ten core is :D

I was only promoting the Titan X because for many GPU compute people I've read complaints of their simulations not fitting in memory. I just wanted to cover that angle for you. If you are perfectly satisfied with 6GB ram, get the two 980 Tis for sure! As I said before, the PSU can handle SLI.
 
You must log in or register to reply here.
Back
Top