CPU for ab-initio DFT calculations

I need to build a new desktop PC, where ab-initio DFT calculation going to be performed. I am searching for a CPU in value range 600 - 1000€. I was thinking about six-core Intel Core i7-6850K or 8-core Intel Core i7-6900K. However Intel Core ix processors are desktop processors so they are optimized for multimedia application and games, I think there are some features I don't need. In case of DFT calculation the most important stuff are matrix-matrix multiplication and FFT. So I was thinking also about Intel Xeon processors. In the same price I can buy 8-12 - core CPU, but with lower frequency (Intel Xeon Processor E5-2630 v4). Now I dont know what to buy.

I cannot find any DFT CPU benchmarks on the web. Can you help me please?

EDIT:

To make it more clear: In value range 600 - 1000€ is better to buy intel i7 or intel xeon for DFT calculations. I am aware of GPU computing and that there are also server solutions, but I am not asking about that.

• Is your question about the computer to buy or about benchmarks for DFT? Have you checked this? – nicoguaro Oct 11 '16 at 15:51
• Why not buying a GPU? If you need only single precision, you can get an excellent GPU for 1000€. – Gil Oct 12 '16 at 19:52
• What kind of FFTs (data types), and what size? Large FFTs are typically memory bound, while dense matrix multiplication (in particular BLAS3 *GEMM) will almost always be compute bound. – njuffa Oct 12 '16 at 22:54
• @nicogauro Thank you for that link but, I was searching for something else. I going to edit question to make it more clear. – Addman Oct 13 '16 at 8:23
• @njuffa Data types are double precision floats. Size of FFT depends on size of the system. Typical FFT dimensions are from 10^3 - 1000^3. In DFT theory, one is calculating one part of energy in real space and the other in furrier space. So you have to travel from one space to another and back all the time. I dont know what is the amount of dense matrix-matrix multiplications, but I think FFTs is more important. – Addman Oct 13 '16 at 8:48

2 Answers

Not aware of any CPU benchmarks for DFT. This might be due to different programs being built to take advantage of different architectures etc. If you are truly looking for pure computational power and do not intend to use the computer in any other way (gaming, movies etc.) then it might be worthwhile to look into buying a server computer.

A quick search in ebay will give you a good idea about the hardware on the market. For example, your price range can get you a HP ProLiant DL585 G7 Server, which has 64-cores and 192 GB of RAM.

On the down side, setting up a computer like this is not a trivial thing, and you will probably not be able to convert it to a gaming computer if you change your mind. It is most likely also going to be loud and produce more heat than a normal PC.

You might also want to check whether or not the program you are using is capable to fully use all the hardware you are about to purchase. 12 core CPU is not going to give you much edge if the program you use can only utilize a couple of them.

• Any reason not going on a GPU? – Gil Oct 12 '16 at 19:44
• No, not really. GPUs occurred to me as well, but I thought I'd limit my answer to CPUs, since the OP asked about them. Also, I've only used GPUs to run MD simulations. – Bdrs Oct 12 '16 at 19:49
• I actually meant to write to Addman. My mistake. – Gil Oct 12 '16 at 19:51
• I is going to only used for DFT. DFT codes are written as parallel code (mpi parallel). Im runnig them on super computer with 1024 cores. – Addman Oct 13 '16 at 9:06
• I don't think GPU with help here as the version of ABINIT he is talking about here does not include GPU parallelization. I recommend you to look at CUDA version of Quantum ESPRESSO that allow you to use GPUS for compputations – Miguel Carvajal Nov 28 '17 at 15:13

I don't know how complex your calculations going to be. If you are planning to use any of the highly parallelized scripts like VASP, I would prefer Xeon processors. Most of the Intel based High Performance Computing (HPC) systems in the world employs Xeon. First task is to confirm the paralleization capabilities of your DFT package. In any case, I recommend you to contact Intel directly. Since it is not so trivial like buying a commercial system for general use, they have a separate section dedicated to handle HPC concerns ( ITP-HPC@intel.com ). I hope this helps.