For pure computational desktops it really depends just *how* hardcore you want. Intel Xeon is the real solution to this; but $2000 is pretty optimistic for a Xeon PC.

My trading computer were Xeon dual quad core 16-32 g of ram. But they are around $5000

Get a fat processor and lots of ram. You probably also want an SSD big enough to store any files that will be read from/written to regularly. Whether you want to prioritize raw gigahertage or more cores (they usually come hand in hand but there's some wiggle room) depends on how exactly you do your computation.

It's pretty easy to build a desktop yourself and a great way to save a lot of money, as most high end pre-built rigs will have more graphics capability than you need and parts are generally cheaper bought individually.

Are you going to be running Linux? How many people at once will be using the machine? How much storage do you need? Building your own or buying?

Also, a lot of work has been done on using GPUs to perform numeric calculations. Do you have any interest in this?

Yeah. The hardcore solution to this is a dual processor motherboard; two Xeon E7 10-18 core CPU's from the 8800 family.

Chuck in 32-64GB DDR4 RAM, and maybe four 512GB SSD's running RAID-0 (or RAID-1 if you're not that brave).

That, of course, is the total extreme for a home desktop though.

Also, out of curiosity, what sort of math do you do?

Thanks a lot, guys :)

I'll look into Intel Xeons, that sounds good!

This is my work computer, and I have to order it through some internal service here at NYU. I don't know how flexible they are when it comes to building

Type of computation: Lots of C++, Matlab (unfortunately). Most of my work right now is behavioral modeling, which means a lot of simulations and optimizations, but not that much in terms of large data files. I am, however, thinking of doing MRI, and those files easily go into 10's of Gigabytes or more. So I'm definitely getting a significant amount of SSD memory.

I have tried working on a GPU before, but I don't know enough about them to make full use of their computational power. Also, for the really heavy stuff, I can move to a computer cluster provided by the university, where I can often get ~100 cores simultaneously to run jobs. They have Intel Xeon E-2690v2 3.0GHz CPUs.

One more note: Almost all the code I write is easily parallelizable - and I know enough to do basic multi-threading. I suppose that means more cores is better than more gigahertz per core, right?

"I suppose that means more cores is better than more gigahertz per core, right?"

This is a very hard question to answer, but, yes, you'll likely want more cores. Of course, faster speeds are good, too.

Do you guys know any good websites for me to do my own research on this? I've looked at the Intel and AMD websites, but they're not that great, and most comparison websites like cnet or pcmag seem to be targeted for gaming pcs.

http://www.tomshardware.com/t/workstations/

Don't touch AMD with a bargepole. Xeon is definitely the way to go for intense computation.

HR knows what he's about. For my 2 cents if you just wanted to keep it simple just get a quality motherboard, an intel i7 CPU, 16GB RAM, and a samsung or Intel SSD drive, and some backup software imaging to external storage.

Keep in mind the best hardware inn the world can be wasted by inefficient software. So your best resource expenditure is on making sure your system is running a minimum of background processes whilst you are doing your number crunching.

With a bit of effort you can strip down Win 7 to a fairly lean and mean thing, especially if you throw Windows Updates overboard. Otherwise linux might be a good idea - but there you really need to know a lot to be sure you are getting the best out of the whole setup.

AMD's fine on a budget, HR.

Make sure it's Win 7 Pro; in 2020 I believe the other versions will no longer receive security updates. Xeon for sure, though.

AMD is fine for gaming on a budget. It's top-end processors have nothing anything like comparable to Xeon though.

A good PC does not a workstation make. Bas will be much better served with Xeon over i7. They are designed to accomplish different things.

I think it's time I take credit for starting yet another great initiative.

Stephee, make a thread titled "Utilize this Thread by Posting your Computer Building Questions Here and Only Here."

Also, Bas, let us know what your Uni says about purchasing computers. I want to know if we get to design a wicked rig

Hmm, wouldn't it be confusing if several people received advice in the same thread?

Yes. It's a joke. Please don't make such a thread.

I knew it was a joke, but then I thought, as I stated: "Hmm".

So, I gotta talk to my advisor to know how much I'm actually allowed to spend.

Regarding software & OS: I mainly use C++ for coding, but I have to use Matlab - it's the de facto standard in neuroscience/psychology, which is what I study. I have some experience with Linux, and I want to get more, so I'm planning to install Linux. That, or dual-boot, but I don't know if that'll reduce performance.

why did nobody recommended a good gpu aswel? those are made for parallel computations...

Because "good GPUs" are quite a bit above his budget, and require a different type of programming, and to quote Bas "I have tried working on a GPU before, but I don't know enough about them to make full use of their computational power."

If you are planning to learn it, I recommend getting a cheap gaming level GPU so you can quickly develop locally, and then perform the actual computations on the Uni HPC or on AWS or somewhere. I can provide some insight into the GPUs if you care about it at all.

Also because if you read the thread y see GPUs have already been discussed

I am actually looking into GPUs. Right now it's a trade-off between performance, money and my own learning curve.

On that note, what do you suggest for GPU programming? I once read a book on CUDA, that seems very powerful but also pretty complicated. I've heard people talk about MPI, but I don't really know how that works...

So I have a quote: I'm allowed to spend $1,800. I could probably go over a little without people getting angry (or I could offer to spend a few of my own dollars on my machine), but $2,000 really is the max.

Is that for just the computer or so you need a monitor, keyboard, etc?

That price would include the monitor, keyboard, mouse etc - although I may be able to find a I'll probably buy a headset and an external drive from my own money, so I can take them with me after I graduate. I'm not going to play video games on this machine, so I don't really care about having a high-resolution screen or a fast mouse/keyboard. I'm mostly going to run code and procrastinate by checking Webdip :)

Understood. All the same, if you're writing a lot of code, its nice to have a good keyboard and monitor. Will you be running Linux?

For this price I think you can make a very nice system. Would you like advice on components?

MONITORS ARE CHEAP

>On that note, what do you suggest for GPU programming?

MPI solutions tend to work with multiple processes, so they are best used when running calculations on a distributed network of many computers. You can use MPI on a single PC, but there are options which are easier to code and are more performant.
For a GPU, using MPI would make sense to sync between different GPUs (or on a distributed network), but really have nothing to do with writing a program that runs on one GPU.

For generic GPU programming, you realistically have only two options - CUDA and OpenCL. Please keep in mind that NVIDIA pays me, so I'm biased.

The showstopper:
If you need the code you write to work on AMD cards, CUDA is out of the question. Not entirely, they have this new thing coming up, but it isn't released yet, and will only work for their $5k+ cards, etc.

If you'll be working on NV cards anyway, CUDA is objectively better supported (we barely update openCL), better documented and you'll find more code and examples for it online. Also, better utility libraries.
Subjectively it may be easier to work with, as it integrates into C++ better than (current) openCL.


If you are not sure whether to go NV or AMD, a few helpers:
NV if you're doing anything related to Neural Networks / Deep learning
NV if you're doing video encoding/decoding
AMD if you are performing a bunch of simple calculations, such as bitcoin mining, crypto stuff, etc
NV if you are writing relatively complex algorithms on the GPU.

AMD cards can perform more raw operations per second than same tier NV cards, but they start getting more and more overhead the more complex your task becomes.