Scientific Computation System Build Advice - Motherboard, GPU, etc.

Toggle sidebar Toggle sidebar
Y

yavinbase

Reputable
Mar 2, 2014
3
0
4,510
My goal is to build (from ~scratch) a system on which I can run parallelizeable scientific simulations (ie, more cores is better). I will be occasionally using the system for gaming. I would like to familiarize myself with overclocking, and being able to temporarily boost my computer's performance during lengthy simulations would be a bonus. Stability is important, and I do not want to have to tweak voltages on a regular basis.

I have done some research and put together a preliminary parts list on pcpartpicker (end of the post). Explanation is required on a few parts. First, the graphics card is from my current computer - I'm not purchasing a new one yet. The operating system and applications will reside on the hard drive, and I will use Sandisk ReadyCache to accelerate boot times and general application performance. The Extreme II will contain any large data files needed by my simulations. These data files obviously need fast I/O, hence the high-performance SSD.

My Questions:

    ■ Will Sandisk ReadyCache automatically move my swap file to the cache? If I wind up having a 64 GB swap file (which is bigger than the capacity of ReadyCache), will the ReadyCache have problems? My hope is that ReadyCache will leave the swap file alone, and I could put the swap file on the Extreme II.
    ■ I have looked at many motherboards, and in particular, I am deciding between P9X79 Pro, P9X79-E WS, Rampage IV Extreme and Rampage IV Black Edition. I tentatively selected the P9X79 Pro because I saw people had overclocked it to 5.0 GHz, and I assume that this means (among other things) that the power is very stable. My question is: Assuming the I/O I need is all there, and assuming that I don't really care about the sound hardware, what difference will I see between the motherboards? Will the answer be different if I need to run 2-4 GPUs in parallel? Are there any new motherboards scheduled/rumored to come out in the next few months that might affect this decision?
    ■ I love the OC panel that comes with the RIVBE. I saw somewhere that it might be sold separately, but I have not found it. Is there any news on this? Are there comparable things that might work for a P9X79 motherboard? In particular, is there something that provides similar functionality to the "normal mode" (ie, displaying progress during POST, displaying voltages, CPU temps and fan speeds, and being able to change between OC profiles quickly)?
    ■ If I need to run double-precision GPU computations, is there a good source of double-precision benchmarks? What about benchmarks of multiple lower-end ones run in SLI or Crossfire?
    ■ I am debating making a custom cooling loop, because I think it would be fun. Are there any considerations which might change my decision?
    ■ Do you have any further comments, concerns or considerations about this build?


Approximate Purchase Date: I am going to build a PC from scratch within the next month or so.

Budget Range: $1500-$1700 after rebates and shipping.

Usage from Most to Least Important:

    ■ Parallelizeable simulations
    ■ Gaming
    ■ Stable Overclocking
    ■ Internet browsing (many tabs)


OS Purchase: I already have an operating system, so I do not need to purchase a new one.

Preferred Website: Newegg, Amazon, TD. No real preference among reputable vendors.

Location: Boulder, CO. I will be going to Microcenter to purchase the CPU.

Parts Preferences: Intel CPU

Overclocking: Yes

SLI or Crossfire: It is possible that GPU computation will lend itself nicely to some (definitely not all) of my computations. If so, then I will probably try out SLI or Crossfire.

Additional Comments: I would like to have a quiet PC.

Why are you upgrading? My old computer does not perform computations quickly enough.



http://pcpartpicker.com/user/yavin/saved/3NNP
PCPartPicker part list / Price breakdown by merchant / Benchmarks

CPU: Intel Core i7-4930K 3.4GHz 6-Core Processor ($499.99 @ Microcenter)
Motherboard: Asus P9X79 PRO ATX LGA2011 Motherboard ($295.98 @ SuperBiiz)
Memory: G.Skill Ripjaws Z Series 16GB (4 x 4GB) DDR3-1866 Memory ($189.99 @ Newegg)
Storage: Sandisk ReadyCache 32GB 2.5" Solid State Disk ($39.99 @ Best Buy)
Storage: Sandisk Extreme II 240GB 2.5" Solid State Disk ($159.98 @ Newegg)
Storage: Western Digital Caviar Blue 1TB 3.5" 7200RPM Internal Hard Drive ($59.98 @ OutletPC)
Video Card: Biostar Radeon HD 4350 512MB Video Card ($55.00 @ Amazon)
Case: Cooler Master HAF 932 Advanced ATX Full Tower Case ($149.99 @ NCIX US)
Power Supply: SeaSonic 760W 80+ Platinum Certified Fully-Modular ATX Power Supply (Purchased For $99.00)
Optical Drive: Pioneer BDR-209DBK Blu-Ray/DVD/CD Writer ($69.98 @ OutletPC)
Total: $1619.88
(Prices include shipping, taxes, and discounts when available.)
(Generated by PCPartPicker 2014-03-02 16:40 EST-0500)

 
Solution
bambiboom
yavinbase,

In my view, a scientific computational /simulation use is incompatible with a high performance gaming use- trying for high performance in both uses will end in disappointment in both uses.

For scientific use- MATLAB for example, for error-free statistical and simulation - precise waveforms, graphs, the best results are obtained with ECC error-correcting RAM, requiring a Xeon CPU, and suitable motherboard. As Xeons are designed for long periods of continuous use they have locked multipliers and can not be overclocked significantly. I think there is some latitude to raise the voltages, but there goes the design reliability. ECC RAM does continuous parity checks and so the latency is lower than for gaming RAM.

As...
Sort by date Sort by votes


Thanks for taking the time to respond. I failed to mention that, for a number of reasons, I do not want to purchase a graphics card right now. I will purchase a new graphics card in a few months, when I know whether to purchase one which is better for scientific computing or for gaming. I think your solution cuts back on the cost of everything else in order to fit a graphics card into the budget.

I also failed to mention that I do not want to significantly compromise performance of the primary function of the machine: scientific computing. I am concerned that your proposed build does this in a number of ways. The one that stands out to me the most is the 2x8 RAM. If I use 2 sticks, I am halving the maximum possible bandwidth. On top of that, the ram clock frequency is 30% lower than the one I originally posted. If I understand correctly, the only reason to use the ram you pointed out is that it costs less and 8 GB sticks will allow me to reach the full potential memory of the motherboard.

There are some things I will reconsider in light of your post. Namely, I may get a cheaper blu-ray writer, and if I can find parts (ram) of comparable performance which are cheaper, I will likely do so.
 
Upvote 0 Downvote
yavinbase,

In my view, a scientific computational /simulation use is incompatible with a high performance gaming use- trying for high performance in both uses will end in disappointment in both uses.

For scientific use- MATLAB for example, for error-free statistical and simulation - precise waveforms, graphs, the best results are obtained with ECC error-correcting RAM, requiring a Xeon CPU, and suitable motherboard. As Xeons are designed for long periods of continuous use they have locked multipliers and can not be overclocked significantly. I think there is some latitude to raise the voltages, but there goes the design reliability. ECC RAM does continuous parity checks and so the latency is lower than for gaming RAM.

As for GPU's, if your software uses viewports, a Quadro or Firepro is necessary to run the special drivers. These cards also have error correction and there can be a substantial benefit with some programs with CUDA acceleration and GPU coprocessing. With CUDA, positional calculations are partly processed in the GPU and this relieves certain CPU computational functions. However, workstation cards use drivers that fully finish every frame instead of higher frame rates, so the gaming experience is somewhat limited. In this respect, it is better to begin with the characteristics of the programs. Again, workstation cards are not fantastic at games, though from the benchmarks, I would think a Quadro K5000 would be respectable, but that is an $1,800 card.

Programs such as MATLAB and simulations are among the most intensive CPU uses and most in need of precision. If your use was #d modeling and rendering, then a system that was alos good at gaming would be possible, but after my own tries with gaming / consumer gear, as soon as my 3D CAD projects became complex, I had severe problems- artifacts and misplaced shadows and particles- signs of computational errors.

That said, if you should decide to opt for a quite performance, scientific priority, very good results are possible within a reasonable budget. Here are some suggestions >

Processor> Intel Xeon Six-Core Processor E5-1650 v2 3.5 / 3.9GHz 0GT/s 12MB LGA 2011 CPU, OEM > $600 (Superbiiz)

CPU Cooling > CORSAIR Hydro Series H80i High Performance Water/Liquid CPU Cooler. 120mm > $89 ( I believe a custom cooling system with reservoir is not necessary)

Motherboard> ASUS P9X79 WS LGA 2011 Intel X79 SATA 6Gb/s USB 3.0 SSI CEB Intel Motherboard with USB BIOS > $370 (This MB is seen in some of the highest performing single CPU workstations)

Motherboard> ASUS Z9PA-U8 ATX Server Motherboard LGA 2011 DDR3 1600/1333/1066 > $278

RAM> 32GB ( 4 X 8 GB) SAMSUNG or Kingston 240-Pin DDR3 SDRAM ECC Registered DDR3 1600 Server Memory Model > about $380-400 (Check tested compatibility with ASUS approved vendors list)(The use of 8GB modules allows an eventual 64GB)

RAM> 16GB ( 2 X 8 GB) SAMSUNG or Kingston 240-Pin DDR3 SDRAM ECC Registered DDR3 1600 Server Memory Model > about $180-200 (Check tested compatibility with ASUS approved vendors list)

HD > SAMSUNG 840 Pro Series MZ-7PD256BW 2.5" 256GB SATA III MLC Internal Solid State Drive (SSD) > $215 (OS, Applications) (I have good results by simply having a partition on the SSD drive for the current working files- very fast loading and saving and less risk of loss than a pure RAMdisk. I also designate any auto-backups to folders on the storage drives that are in RAID 1 mirroring)

HD Seagate Constellation ES.3 ST1000NM0033 1TB 7200 RPM 128MB Cache SATA 6.0Gb/s 3.5" Enterprise Internal Hard Drive Bare Drive > $117 (Add a 2nd later in RAID 1) (Note the 128MB cache and 1.2M MTBF)

HD WD Re WD1003FBYZ 1TB 7200 RPM 64MB Cache SATA 6.0Gb/s 3.5" Datacenter Capacity Internal Hard Drive Bare Drive > $110

_____________________________________________________________________

Cheers,

BambiBoom

HP z420 (2014) > Xeon E5-1620 quad core @ 3.6 / 3.8GHz > 24GB ECC 1600 RAM > Quadro 4000 (2GB)> Samsung 840 SSD 250GB /Western Digital WD1003FZEX 1TB> M-Audio 192 sound card > AE3000 USB WiFi > HP 2711X, 27" 1920 X 1080 > Windows 7 Ultimate 64 > Autodesk Building Design Suite, Inventor Pro, Solidworks, Adobe CS MC, Corel Technical Design Sketchup Pro, WordP Office, MS Office Pro [Passmark system rating = 3815, 2D= 767 / 3D=2044]

Dell Precision T5400 (2008) > 2X Xeon X5460 quad core @3.16GHz > 16GB ECC 667> Quadro FX 4800 (1.5GB) > WD RE4 500GB / Seagate Barracuda 500GB > M-Audio 2496 Sound Card / Linksys 600N WiFi > Windows 7 Ultimate 64-bit > (earlier versions) AutoCad, Revit, Solidworks, Sketchup Pro, Corel Technical Designer, Adobe CS MC, WordP Office, MS Office Pro [Passmark system rating = 1859, 2D= 512 / 3D=1097]

 
Upvote 0 Downvote
Solution


This is a very interesting post. Thank you for the insight. This gives me much to think about before ordering components.

To ensure I understand correctly: The main benefit of using the workstation components is that they support ECC RAM, which has lower latency and, as its name suggests, minimizes errors. Has there been research done on the amount of error accumulated in today's gaming ram as compared to normal ram? Wikipedia suggests that in 2007-2009:

The bit error rate ranged] from 10^−10 to 10^−17 error/bit·h, roughly one bit error, per hour, per gigabyte of memory to one bit error, per millennium, per gigabyte of memory.
Click to expand...

Is ECC RAM likely to make a visible difference if my simulations run for less than an hour a few times a day? It seems that it is not, but I obviously do not have experience in this area.
 
Upvote 0 Downvote
yavinbase,

The statistical significance of error-correcting RAM seems at first to be insignificant, but then consider the sheer number of bytes over a very short time. A friend of mine is doing flight simulations using MATLAB, applying low pass filters to in effect smooth anomalous peaks from waveform data (von Karman). I was interested in trying to work on the MATLAB portion to learn it, but I can't write the necessary algorithms- very complex- and the filtering uses 1st, 2nd, 3rd, and 4th order passes. My friend was explaining the importance of error free simulations as one misplaced bit accumulates errors at an astounding rate in the filtering and is multiplied in the predictive analyses. The latency of ECC is actually longer than from non-ECC because of the parity checks, a reason why scientific systems are not optimal for gaming, but in my view, the error correction is necessary for precision. This is true in the visual side as well as pure computational as ECC graphics cards have more accurate- and artifact free shadows, particles, reflections, color gradients, and so on. Workstation cards can in some applications (e.g., Solidworks) produce 128x anti-aliasing ( a step / angle is different data than a curve) and if you think about what that represents in terms of an accurate waveforms, it can be significant in analysis of, for example, particle annihilation and aerodynamics. I would strongly advise using an ECC system.

As computers have become more powerful ,so have they become more specialized and workstation and gaming system continue to diverge in hardware. I have never played a video game, and while I understand the need for high frame rates at high resolution and the hardware that does it, I don't have the experiential history to really optimize a system for gaming. Likewise, those who don't use workstation applications at a high level will be missing some important aspects that will skew hardware choices. Content creation and content consumption systems- above a certain performance level- are becoming more and more divergent.

Cheers,

BambiBoom

 
Upvote 0 Downvote
You must log in or register to reply here.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom