Need a good build for scientific computing. Please advise.

[college_student]

Hi all,

I need to build a computer to use ONLY for scientific computing. It will be used to run code for research in image and video analysis. Here are answers to the thread:

1. Scientific computing. Writing code and analyzing images, video, etc. NO GAMING! I don't want a server. I need A LOT of RAM. Likely max out the RAM -- 64GB if I'm correct.

2. $2000 - $2500

3. USA

4. USA

5. Intel. Buying everything from Newegg.

6. Monitor, Keyboard, Mouse, and OS

7. Probably over clock at some point. Nothing crazy.

8. 1080p - Dual Monitors

9. ASAP

This is what I'm thinking to get:

CPU - 3930K ($569.99)
HDD - WD Black 1 TB ($89.99)
SSD - Samsung 840 Pro 256GB ($249.99)
RAM - 64GB Corsair DDR3 ($399.99)
Mobo - ASRock X79 Extreme6 ($219.99)
GPU - HD 7970 GHz Edition 6GB ($579.99)
PSU - SeaSonic USA X-1050 Fully-Modular 80PLUS GOLD ($199.99)
Case - Cooler Master HAF 932 Adv RX-932-KKN5-GP ($139.99)
Cooler - ZALMAN 135mm Long CPU Cooler ($76.99)
Total: $2526.91

[T_Yamamoto]

I would go with a duel xeon build if you're going to do coding.

[Vincent]

Did you take a look at this guide?

http://www.anandtech.com/show/6492/w...n-buyers-guide

[T_Yamamoto]

Originally Posted by Vincent

Did you take a look at this guide?

http://www.anandtech.com/show/6492/w...n-buyers-guide

It would be better to go with a 2P intel or 4P AMD rig for scientific computing

[Sleepingforest]

I would buy from an OEM if reliability and downtime is a big issue. You pay a premium for next-day support and minimized downtime. Otherwise, go eith as many cores as you can.

[college_student]

Originally Posted by Vincent

Did you take a look at this guide?

http://www.anandtech.com/show/6492/w...n-buyers-guide

No, but now that I look at it, they seemed pick the same CPU and more less the same parts (except for GPU). The GPU I picked has 6GB RAM, which makes it quite good for my GPU computing.

Originally Posted by T_Yamamoto

It would be better to go with a 2P intel or 4P AMD rig for scientific computing

Originally Posted by T_Yamamoto

I would go with a duel xeon build if you're going to do coding.

While I agree that two server grade processors (Xeon, etc) would be nice, they are unfortunately outside my budget ($2.5K max).

Any critiques on my specific build?

[DSF]

If you mean that Xeon processors are outside your budget, you're mistaken.

Xeon 1230 V2:
http://www.newegg.com/Product/Produc...82E16819117286

3.3 Ghz Ivy Bridge with Hyperthreading for $240.

[college_student]

Originally Posted by DSF

If you mean that Xeon processors are outside your budget, you're mistaken.

Xeon 1230 V2:
http://www.newegg.com/Product/Produc...82E16819117286

3.3 Ghz Ivy Bridge with Hyperthreading for $240.

Hi DSF,

So what would you suggest then? Two of those? Also, what is your opinion on the 1230V2 vs. 3930K?

Do you have a Mobo in mind that can take 2X those and 64GB RAM?

From what I understand, a server CPU means I need a server Mobo, and therefore often I need the Fully-bufferred ECC RAM, etc; all in all these components tend to cost more than the 3930K route (in a price vs. performance sense).

Thanks!

[bononos]

Originally Posted by college_student

...
From what I understand, a server CPU means I need a server Mobo, and therefore often I need the Fully-bufferred ECC RAM, etc; all in all these components tend to cost more than the 3930K route (in a price vs. performance sense).

Thanks!

I didn't check the link but there are socket 1155 Xeons which don't need ecc ram.

[DSF]

However, I'm not sure there are any dual-socket LGA1155 boards, that would be the issue. I hadn't realized that.

When you mention price/performance though, I'm still not sure a 3930K gives you that. Does the processing time affect income for you? If not, you could save significant cash with an LGA1155 setup. If processing time is tied to your bottom line, then maybe LGA2011 makes sense.

[Torn Mind]

Originally Posted by DSF

However, I'm not sure there are any dual-socket LGA1155 boards, that would be the issue. I hadn't realized that.

When you mention price/performance though, I'm still not sure a 3930K gives you that. Does the processing time affect income for you? If not, you could save significant cash with an LGA1155 setup. If processing time is tied to your bottom line, then maybe LGA2011 makes sense.

I believe it is impossible on LGA 1155 due to not having a second qpi link?

Or at least that is what I found at this following site:
http://techreport.com/forums/viewtopic.php?f=36&t=80996

[college_student]

Originally Posted by DSF

However, I'm not sure there are any dual-socket LGA1155 boards, that would be the issue. I hadn't realized that.

When you mention price/performance though, I'm still not sure a 3930K gives you that. Does the processing time affect income for you? If not, you could save significant cash with an LGA1155 setup. If processing time is tied to your bottom line, then maybe LGA2011 makes sense.

Right, I didn't mean that 3930K maxes out the price/performance ratio. I meant for my needs and budget, it might be the best (idk, I'm speculating).

If I understand the second part of your post correctly, then no, processing time doesn't affect my income. I need the 64GB of RAM because the images and videos I deal with are HUGE and eat up RAM like crazy. As such, the only non-server Mobo's that I found (in my budget) that accept 64GB of RAM are the LGA2011 boards. And as far as processing time, I would prefer if I could run my simulations, etc quickly, so that I could move on to the next step quicker. This is for R&D work, so while time isn't linked to income, it does affect how long we have to wait before we move to the next step. I hope this makes sense. Please tell me if it doesn't.

I should note that price is not an issue as long as I stay below $2K-$2.5K. I want to get the best bang for the buck for $2.5K, not necessarily the best bang for the buck in general. In summary, what is the best performance that can be had for $2.5K?

Originally Posted by Torn Mind

I believe it is impossible on LGA 1155 due to not having a second qpi link?

Or at least that is what I found at this following site:
http://techreport.com/forums/viewtopic.php?f=36&t=80996

Hmm interesting!

[mfenn]

Yes, 1155 cannot do dual CPU because it doesn't have ANY QPI links, only DMI to the chipset and PCIe to the slots.

Originally Posted by college_student

This is what I'm thinking to get:

CPU - 3930K ($569.99)
HDD - WD Black 1 TB ($89.99)
SSD - Samsung 840 Pro 256GB ($249.99)
RAM - 64GB Corsair DDR3 ($399.99)
Mobo - ASRock X79 Extreme6 ($219.99)
GPU - HD 7970 GHz Edition 6GB ($579.99)
PSU - SeaSonic USA X-1050 Fully-Modular 80PLUS GOLD ($199.99)
Case - Cooler Master HAF 932 Adv RX-932-KKN5-GP ($139.99)
Cooler - ZALMAN 135mm Long CPU Cooler ($76.99)
Total: $2526.91

This is a gaming rig, not a scientific computing workstation. To really give you good advice, we need to know more about your code.

- What is the goal?
- Are you writing it from scratch or modifying an existing codebase?
- What operating system?
- Does it use any GPU acceleration or are you planning to add any?
- What does the I/O profile look like?

[Torn Mind]

Single GPU systems don't need more than a 500 watt PSU, 600 watts tops, as long as the PSU can deliver what it is specified to deliver.

[mv2devnull]

Originally Posted by mfenn

To really give you good advice, we need to know more about your code.

Indeed. What is "scientific computing" in this case?

[college_student]

Originally Posted by mfenn

Yes, 1155 cannot do dual CPU because it doesn't have ANY QPI links, only DMI to the chipset and PCIe to the slots.



This is a gaming rig, not a scientific computing workstation. To really give you good advice, we need to know more about your code.

- What is the goal?
- Are you writing it from scratch or modifying an existing codebase?
- What operating system?
- Does it use any GPU acceleration or are you planning to add any?
- What does the I/O profile look like?

Hi mfenn,

This machine is going to be used to write software that analyzes extremely large image cubes and videos. A lot of the software is going to be written in common programming languages: C++, OpenCV, etc. I plan to use the GPU for GPU acceleration; I picked this one because it had 6GB of memory so it could store some very large images onto itself. I'm looking into OpenCL.

A lot of the software already exists, but a lot of it will be written from scratch. The goal is for the machine to be able to handle the data I give it and be able to do the tasks we need in a reasonable amount of time. You can think of it like this: this machine will constantly be doing complex operations on extremely large, multi-dimensional matrices.

I/O? Typically we write scripts that fetch the relevant data from a server or a drive, load it into memory, perform operations on it, output the result (sometimes numerical, sometimes a plot, sometimes images/video, etc).

I will use Linux.

Thanks!

[mv2devnull]

Originally Posted by college_student

GPU - HD 7970 GHz Edition 6GB
I will use Linux.

AMD drivers in Linux seems to be a controversial topic, but the raw power of 7970 is undeniable.


Images, but not quite what the Belgians did back in 2008?
http://fastra.ua.ac.be/en/index.html

[mfenn]

Originally Posted by college_student

Hi mfenn,

This machine is going to be used to write software that analyzes extremely large image cubes and videos. A lot of the software is going to be written in common programming languages: C++, OpenCV, etc. I plan to use the GPU for GPU acceleration; I picked this one because it had 6GB of memory so it could store some very large images onto itself. I'm looking into OpenCL.

A lot of the software already exists, but a lot of it will be written from scratch. The goal is for the machine to be able to handle the data I give it and be able to do the tasks we need in a reasonable amount of time. You can think of it like this: this machine will constantly be doing complex operations on extremely large, multi-dimensional matrices.

I/O? Typically we write scripts that fetch the relevant data from a server or a drive, load it into memory, perform operations on it, output the result (sometimes numerical, sometimes a plot, sometimes images/video, etc).

I will use Linux.

Thanks!

That's a good start and it gives me some idea of what you're doing. There's some critical quantitative data that's missing though. What you say "large", what do you mean? Large to some people is a gigabyte, to others it's 100 gigabytes. Another question, do you care about double-precision or is single enough for you? Most image processing algorithms only use single-precision, but it doesn't hurt to check.

I should have been explicit that I was looking for both quantitative and qualitative data, sorry about that.

Anyway, since you will be using/extending existing code, writing your own code (presumably you have not done GPU programming before), and will be using Linux; I highly recommend that you use an Nvidia GPU. CUDA is much easier to learn than OpenCL, you're much more likely to run into codes that use CUDA over OpenCL, and there are better support libraries for CUDA (you'll probably be interested in cuBLAS). Furthermore, if your code scales up enough to run on a real cluster, you will be much more likely to find an Nvidia cluster than an AMD one (like 20:1 ratio).