When I first saw that Intel was going to support Xeon Phi on Windows I asked myself –why would they do that? Surely no one is interested in that! It's an embedded Linux device after all and it's intended for high performance computing use. No one seriously does HPC on Windows!
Then people started asking me about Windows and Phi. As an example, I talked with someone who was using some Windows based software for important work and their job runs were taking a week or longer to complete, and they were using good machines to run the jobs on too. They were trying to convince the software developers to port the code to Phi to hopefully speed things up a bit. Does Phi really work on Windows? I decided to try setting up the beta drivers for Windows to see if it really did work. Sure enough, it went OK. I had to add some things like an ssh and scp client (Putty and WinSCP) and setup ssh keys and other things you may not "normally" associate with Windows but it wasn't really that bad to setup. I wrote a blog post about it…
Note:The WHQL driver was released on Sept 27. Check out MPSS at Intel.
Questions about Phi and Windows kept coming in… Then something else happened that surprised me. I was sitting in on an Intel webinar about the Xeon Phi and they put up the following poll question;
What host operating systems have your customers expressed interest for using with Intel® Xeon PhiTM products? a. Red Hat Enterprise Linux* b. SuSE Linux Enterprise Server* c. Microsoft Windows* (Server, Win 7, Win 8) d. Other
87% of the responses were c.! Microsoft Windows!
I was stunned, and so was the guy doing the webinar! It was really unexpected.
Later I had another customer asking me about Phi on Windows. They had some critical software they were using and had good relations with the software developers who had expressed interest in possibly porting to Phi (they had ruled out Tesla and CUDA for various reasons). … Then it hit me… The Phi could be a big hit on Windows because users are desperate for performance!
With Linux parallel codes get done using SMP threads and/or distributed parallelism with MPI routinely since nearly all "super-computers" are just Linux clusters of various node counts. (They are relatively easy to build and Linux is free). So there are lots of Linux codes that scale across cluster nodes to hundreds or thousands of processors. HPC on Linux is common. On the Windows side, it's nearly all single box SMP threads for parallelism, since doing Windows clusters is difficult and expensive. HPC on Windows is uncommon… but there is some very nice and relatively easy to use Windows software that people are very productive with, and depend on! So the promise of multi-teraFLOP performance on a Windows platform is REALLY appealing to them. It will be interesting to see how this all develops over the next year or so.
I'll mess around with Windows and Phi some more and keep you posted.
Happy computing! –dbk