what is the best for HPC computing?

[BishopWolf]

I wonder what is the best for HPC computing, an Intel Xeon Phi 5110P or a Nvidia Titan Z on a $8000 budget. What did you recommend me.

 

[kanewolf]

Maybe neither. Maybe a Tesla. Maybe a pair of Xeons. Maybe a cluster of low power pizza boxes. HPC is too broad a category to provide an answer.

Are you writing scratch C++ code? Using MATLAB? Some other application? Do you know your application is CPU bound? Could you keep a Phi fed with data? Will your application support a co-processor (Tesla or Phi)?

You will need to characterize the problem before any answer can be provided.

 

[Eximo]

Depends on the software really. I couldn't just randomly tell you to pick between two architectures without knowing a little about what it is you are trying to do.

If you need a whole bunch of x86 cores working in parallel, the Xeon Phi is the way to go. Titan Z has two very powerful GPUs on a single card, but is sort of a hybrid gaming/professional card. Tesla cards are the same architecture. If your software supports CUDA or uses OpenGL then maybe these are the better choice.

 

[BishopWolf]

HPC is too broad a category to provide an answer.
Are you writing scratch C++ code? Using MATLAB? Some other application? Do you know your application is CPU bound? Could you keep a Phi fed with data? Will your application support a co-processor (Tesla or Phi)?
You will need to characterize the problem before any answer can be provided.

I am writing C++ code (https://github.com/BishopWolf/Gate) and yes I can feed a Phi with a bunch of data, but my libraries also support CUDA, so I wonder what path gives me better results. In my page you will find a description of the code and the problem.

 

[Eximo]

Well that is a bit tricky. Can certainly get a lot of CUDA cores on a Tesla, 4996 with the K80. Xeon Phi only have 61 cores on the big cards while the lowliest Tesla has 2496.

If you were to clock your CPU lower and execute your compiler on a single core, and then compare that to a single Kepler CUDA core (GTX 600-700) you might have a baseline. I don't have access to any Teslas to test it out myself (we have a few, but they are pretty much in use 24/7)

 

[kanewolf]

After looking at the http://www.opengatecollaboration.org/ web page, I think I would recommend CUDA over Phi for the main reason that there are more users and you will probably have more developers to interact with. The K80 is supposed to be top end when available ($5K at NewEgg but not available). It is passively cooled so you will need a chassis that is intended to run it. Something like a dell R730.

 

[kanewolf]

One way to test your performance would be to use AWS with GPUs -- https://aws.amazon.com/articles/7249489223918169