Help with building a number cruncher

Associate
Joined
2 May 2013
Posts
17
Hi all,

You seem like a knowledgeable forum - and I was hoping you could help me out.

I'm working on some research, that research requires doing a lot of number crunching on ~55GB of data. To make this process as quick as possible, all that data should be in memory. A vast majority of the work it's doing is on floating point numbers. This'll be an iterative process running over several weeks.

I've built a few machines before, although only AMD. I thinking an Intel CPU might be better for this job.

Graphics are not important, and it'll only be dumping a few GB of results to disk every 24 hours, so I'm not that fussed about the drive either. As an afterthought - this machine might eventually become a development workstation once it's done number crunching.

I'd like to spend as little as possible (doesn't everyone), ideally <£1000. Putting together a few pieces as a starting point (And I'm already over budget):

Intel Core i7-3820 3.60GHz (Sandybridge-E) Socket LGA2011 Processor - Retail
2x Corsair Vengeance RED 32GB (4x8GB) DDR3 PC3-14900C10 1866MHz Dual/Quad Channel Kit (CMZ32GX3M4X1866C10R)
Asus P9X79 Intel X79 (Socket 2011) DDR3 Motherboard
Western Digital Caviar Green 1TB SATA 6Gb/s 64MB Cache - OEM (WD10EZRX) **SINGLE PLATTER** HDD
Cooler Master Elite Power 500W Power Supply
Xigmatek Asgard Midi Tower Case - Black
Sapphire HD 6450 HM 512MB GDDR3 PCI-Express Graphics Card (11190-04-20G)
Microsoft Windows 7 Professional SP1 64-Bit - OEM (FQC-04649)
£1,159.91

Thanks in advance!
 
Thanks for the feedback.

@CinderzFTW - I'm pretty certain that replacing the 64GB RAM with an SSD would slow it down massively. I know SSDs are quick, but not that quick. Especially when being read and written by multiple threads.

@JonJ678 - It's crunching the data to produce and adapt a decision tree. All inputs are floats. The reason I'm going with windows is because a majority of the libraries that handle the data are already written in C# + that's the language I know best.

Multiple machines sounds like an interesting idea, especially as I could start with one and expand as needed.

Anyone got any ideas on recommended hardware? Any adaptations to what I've got listed above?

Cheers.
 
I'm re-evaluating my build. Any thoughts on the i7 3930K over the 3820? Less speed, 50% more threads.

If I have 12 threads making memory requests, is the RAM likely to be a bottleneck? Or am I likely to be able to squeeze, say a 25% performance boost?

I know this is all code dependent as well as hardware dependant. Advice would be appreciated all the same.
 
Thanks for that.

@deFiniLoGy - Too many acronyms for a novice. R4E? IB-E? SB-E? Sorry, stuff I should probably know. So you reckon it'd be worth me splashing out on Kingston HyperX Beast 32GB (4x8GB) PC3-19200C11 2400MHz?

@MonsterMoshi I have 2 concerns with your build. You've downgraded to a slower RAM. Any reason other than price saving? Also, I'd need more hard drive space for dumping results, and don't really need the speed of an SSD, as once everything is in RAM it shouldn't have to touch the drive for 24 hours. Nice idea with the OEM CPU and chunky cooler. :-)

Anyone got any experience of the P9X79 in it's various forms? Any other motherboard I should be considering?
 
Going off on a tangent here. If I was to go for the 3820, is that definitely worth the extra money over the AMD FX-8350? Benchmarks generally seem to put the two pretty much on a par. I can't find and gflops numbers anywhere though - which is what I'd really be interested in. Anyone know where I could find accurate comparisons of those? What could I get them to if OCed?

Cheers.
 
Thanks, the second one is very interesting, especially the FPU tests. Summed up nicely in this paragraph.

The FX-8350 takes the top spot in the CPU Hash test, not surprising given the relatively strong performance of the AMD processors in this integer-focused benchmark. The more FPU-intensive fractal tests are a very different story, with the Sandy and Ivy Bridge-based chips topping the charts.
Sounds like the Intel chip wins.
 
The 3930k I mentioned earlier performs almost the same as the 3960x in those benches. If you run the two @the same clock speed it is hard to tell them apart.
The 3930k is a serious consideration, but having seen the tech report link that Idleman posted - the 3930k and the 3820 are pretty much tied on the FPU tests, both are annihilated by the 3960x. I'm not sure exactly what those tests incorporate, but that sounds like the most suitable test for me to compare with.

So having seen that I'm tempted to stick with the 3820.

Edit: Sorry, I wrote a load of rubbish. Just noticed the 3930k isn't mentioned in that test, it's the 3770K.

Edit: Quick bit of research. For those that are interested - Looks like the Julia FPU test runs with a 32bit FP datatype. (Most suitable for me). The Mandel test runs with 64bit FP datatypes.
 
Last edited:
I'm looking at using the 3930K overclocked with the Corsair Hydro H55

I've never used any form of liquid cooler before. Is this a good idea? Any other tips? I guess I should have an extra case fan to help move air through the case?

Thanks.
 
if your doing a lot of floating point calcuations, have you through about using Cuda?
I have thought about using Cuda. It's something I *might* look into in the future, but the vast majority of the code is already written to run on a standard CPU.

I know there are a few Cuda libraries for C# these days, and I'm hoping to have a look at them in the not too distant future.
 
@!bluetonic! - Thanks for the links, I'll take a look.

@JonJ678 - You sound like you know what you're talking about. Due to all the data being stored in a tree, I suspect I'll be hitting the RAM fairly hard as there will be a lack of cache locality, so I might just save myself a couple of hundred quid.

I need new hardware for several reasons, firstly, all my current hardware is a couple of years old now. Secondly, It's all in use, so I don't have anything spare to do the number crunching for weeks on end. The code is already very heavily optimised. Everything has been written for speed from the start.

Although.... is it worth looking at the 3930K over the 3820 due to ease of overclocking, or is this not a serious consideration? I'm a beginner when it comes to overclocking machines.

Thanks
 
Ok, I think I'm getting there. The Sabertooth is a fair bit more expensive, but it sounds like it'll handle the heavy loads a bit better.

So, back to the CPU:
JonJ678 made a pretty compelling case for the 3820 over the 3930K... but is the 3930K a better bet for overclocking? Or will I be able to overclock the 3820 to a decent level with the Sabertooth X79?
 
Getting Expensive

Ok, well I'm ordering tonight. Last shouts if anyone thinks there is something I should change!
Asus X79 Sabertooth Intel X79 (Socket 2011) DDR3 Motherboard
Intel Core i7-3820 3.60GHz
Corsair Vengeance RED 32GB (4x8GB) DDR3 PC3-14900C10 1866MHz x2
SanDisk Extreme SSD 240GB 2.5" SATA 6Gb/s Solid State Hard Drive (So I can use it as a work PC when not number crunching)
Microsoft Windows 7 Professional SP1 64-Bit - OEM (FQC-04649)
MSI GeForce GTX 650 Black Knight OC 1024MB (CUDA enabled for experimenting in the future)
Alpenföhn K2 Mount Doom CPU Cooler
Western Digital Caviar Green 1TB SATA 6Gb/s 64MB Cache (For results dumping)
OCZ ZS Series 550W '80 Plus Bronze' (This seems to be recommended)
Xigmatek Asgard Midi Tower Case - Black
£1,506.88

The price seems to be slowly increasing
 
Why are you going for the £270 Sabertooth over something like these for example?

YOUR BASKET
1 x Gigabyte X79-UP4 Intel X79 (Socket 2011) DDR3 Motherboard £209.99
1 x Asus P9X79 Intel X79 (Socket 2011) DDR3 Motherboard £190.99
That decision was based on what was said earlier about the P9X79
I own this board and wouldn't recommend it for doing lots of high intensity work especially if you plan on overclocking. At stock clocks when folding with my 3820 the VRM heatsink gets hot enough to trigger the human "take your skin of the burney thing" reaction. If you check out the X79 Sabertooth (pretty much the same board with better cooling and a weird colour scheme) it has the same VRM heatsink but connected via a heatpipe to a second heatink with optional fan.
It sounds like you think it's an incorrect choice? Do you have experience with any/all of these boards? Thanks.
 
Thanks, I'm always up for saving a few pennies. Any ideas if I'd need low profile ram on that board with a big cooler?
 
Back
Top Bottom