I am planning to put together a system for number crunching. I would say 90% Matlab which will be running on Windows 7. The types of algorithms I run are mainly for solving high-dimensional fixed point problems and for doing large scale optimization (use the optimization toolbox frequently). Most of the time I have to write my own code and due to the iterative nature it unfortunately relies on a lot of for/while loops which can not be parallelized. I would like to have the option for overclocking in the future as well but that is slightly less important. My main goal is speed and stability (programs can run for a few days at a time), secondly budget (I am trying to stay around $1000) and lastly energy efficiency. Here is what I am thinking about
CPU: Intel i7 960 - I def want to go i7 and the 960 is the highest in my price range.
Mobo: ASUS sabertooth X58 - I do not know much about hardware, I have seen good reviews on this one though and the price seems reasonable.
RAM: 3x4gb Corsair DDR3 1600. I know the i7 is setup for 1066, but I would like to have the option of overclocking in the future. How much will a lower cas latency help, say 7 vs 9?
Hard Drive: hdd 7200 rpm 500 gb. I don't really need more than 500gb on this unit. The types of work I will do do not read/write data off/to the hard drive frequently so I do not have much use for ssd or a 10000 rpm hdd either. I will also do some work with Matlab and fortran on linux but much less. That means I will either partition the drive or install two separate drives. Any recs on which is more convenient?
GPU: nVIDIA HD GeForce 8400??: I do not know much about what I need here. I def want nVIDIA to utilize the CUDA with Matlab though. However, I do not need to go overboard either as I have no plans of much parallel processing as my algorithms can not be parallelized to begin with. However, I do utilize some CUDA functions and it would be nice to get a little extra kick if possible. Is it possible to get some extra speed here with out spending a lot on the GPU?
PSU: I am thinking anywhere from 600-750w. What are some good brands here? Do I need 750 if I plan to overclock to 4 Ghz?
Cooling: Obviously I need a cpu cooler. Should I go with copper pipes or standard? Again, if I decide to overclock what will I need. Do I need a hard drive fan? What is a good case that allows good cooling, and what cooling fans should I get?
The rest, optical drive, sound card, network card, etc I will figure out myself as they are not relevant for computing speed.
Mattlab lol; if you aren't gaming you don't need high end software.
Here's some advice for your pocketbook anyways: buy a i7-2600k and Z68 LGA 1155 motherboard. It's the best CPU out there and it's ~$25 more than the 960.
CAS latency helps but don't worry about it if you don't know what you're doing. Anything DDR3 1333 and higher is more than sufficient.
You won't notice a big difference in mechanical drivers compared to SSD which are still overpriced but could be useful for your purposes; they are faster loading data to RAM for programs.
CUDA is just marketing hype; Intel has for superior instruction sets for their new Sandy Bridge (like the 2600k) processors. Even if mattlab uses CUDA you'll have to get something much faster than the 8400GS (something $90) to take advantage of it.
Go with a high quality PSU; don't listen to the ignorant hype on these boards. Corsair is extremely overrated; XFX and Antec make better units for cheaper. Anything 80 Plus will be sufficient from a reputable (sometimes even disreputable) brand. Overclocking the CPU increases the power draw of your system like 10-20 watts. Video cards suck the most power. For your puroses go with a 500W Antec.
You will only need a separate cooler for your CPU if you overclock.
Thanks for the quick reply ignoranceabound. By the way, I am not sure what high end software you are talking about, Matlab is purely for scientific computing, not gaming. I actually just did a search the 2600k and the folks at Mathworks (makers of Matlab) did some tests and they found that the 2600k is %10 faster so thanks a lot for that rec! Just a few comments/ followups.
Lets say I do use some CUDA commands. Then which graphics card would you suggest? I don't mind spending $90-$100 on a gpu if I get some mileage out of it.
As far as cooling goes, I will need at least one for the cpu even without over clocking. So My question is whether I should upgrade to copper pipes or not.
GeForce run faster than anything else, cause they have the same number of cores, but run at a faster clock speed than Tesla. However, Tesla gives you more memory, ECC memory, and lower-power consumption.
For GPU computing in Matlab, don't mess around with Parallel computing toolbox, which is full of hot air. You're much better off going straight to Jacket (google for it), which is a really fantastic implementation.
Ok. I think I close to squared away with hardware. Now I just need to find someone to build it. I don't want to spend my time putting pieces together, plus I am not experienced so I don't want to make a mistake either.
What do you guys recommend for barebones units? Cyberpowerpc has some good prices but I read terrible reviews about customer service. Others include allpczone which has great reviews, but on their ebay listings they don't appear to have options for the hardware that I want. What are some other suggestions?