GPU vs Phi: advice needed

[yaximik]

A while ago I asked a general question about practical usability of GPU- and Phi- based platforms for multithread-enabled bioinformatics applications. I am aware of previously held discussion here. My understanding that there was not much experience accumulated yet with use of these systems to outline the differences in capabilities. I wonder if there is more of such experience now. Basically I have choice to obtain a RHEL 6 server equipped with 2 8-core Xeon E5 2690 and 4 either K20x GPU or 60-core Xeon Phi coprocessors. The price is about the same. Microway engineers advise to go with GPU as there are more applications developed.
As I am not programmer, rather a self-educating end user, my question is which configuration will be more usable in the current universe of bioinformatis applications without code modifications? Please forgive if my questions sound naive. For example, very likely I will not be able to take advantage of thousands of cores with the GPU configuration when using GNU parallel or multithreading in BWA or blast+ without significant investment in code modification. But what about Phi - can I take immediate advantage of 240 cores with the Phi configuration?

[GenoMax]

We had a recent discussion on this with input from Rick Westerman: http://seqanswers.com/forums/showthread.php?t=20931

I/O and memory needs of most NGS applications make use of either the GPU/Phi impractical at this time. Unless you have a use case identified for either the GPU/Phi, you may be better off putting that money towards upgrade of some other components.

[ymc]

While GPU-based NGS applications are not that many, Phi-based NGS applications are basically nil (Let me know if you know of one). soap3-dp is a GPU-based application that you can use to replace bwa.

I think you will only pick Phi if you have some floating point application that is based on some old code optimized for x86. There aren't many such application for NGS. Therefore it makes no sense to use Phi at all.

[ymc]

We had a recent discussion on this with input from Rick Westerman: http://seqanswers.com/forums/showthread.php?t=20931

I/O and memory needs of most NGS applications make use of either the GPU/Phi impractical at this time. Unless you have a use case identified for either the GPU/Phi, you may be better off putting that money towards upgrade of some other components.

If you have some application that make use of floating point calculations a lot, then the value of GPU will show. For example, I have a CUDA program that calculates pairwise familial relationships between my samples.

Another way of thinking is that using GPU can make your system more compact. For example, you can run soap3-dp on a GPU and then do other stuff with your CPUs. In this sense, your computations per one computer case volume will be higher than without GPU.

[yaximik]

I missed the most recent discussion, thanks for the link, very useful. Also thanks a lot for other useful inputs.