Xeon Phi hardware is starting to trickle into used markets for… affordable?.. prices.
As with any Intel processor, they are compatible with OpenCL, meaning that, at least hypothetically, they would work with Blender and other render programs. I’m looking at this as a possible render accelerator that does all of the things that GPU rendering can’t do.
Intel’s website is surprisingly uninformative about Phi. Does anyone here know anything vis-a-vis Blender?
Keep in mind, Cycles is always in “GPU mode” when running on OpenCL devices, even if the targeted device is actually a CPU. It basically acts as though your CPU or the Phi board or anything else were additional graphics cards. So volumes/SSS still wouldn’t work, at least not in an off-the-shelf build. (SSS can be enabled at build time, but that works for newer Nvidia cards too).
Unlike a GPU, however, the Intel® Xeon Phi™ coprocessor also features:
Intel® architecture compatibility – no need to recode your entire problem or master new tools and programming models
A single programming model for all your code – parallel optimization techniques simultaneously deliver performance for both Intel® Xeon® processor and Intel® Xeon Phi™ coprocessor
Flexible execution models – beyond basic offload, you can also run your application natively on the coprocessor or symmetrically on both host processor and coprocessor
Powers the world’s #1 supercomputer – 48,000 Intel® Xeon Phi™ coprocessors power the world’s fastest supercomputer in concert with Intel® Xeon® processors[SUP]5[/SUP]
I think that the big issue is that Blender would still have to be re-coded to correctly interface with the Phi hardware. I’ve also read that there are issues with what software actually works very well. The software needs to be able to handle an enormous number of threads and also take advantage of Phi’s 512-bit vector units, otherwise performance will be disappointing.
That would be a great idea for Intel to attack a market that the other two big players are pretty much ignoring. I mean, for Pete’s sake, AMD’s drivers are unbelievably broken.
I just read from the Wikipedia page, that the dream of Phi is still a ways away.
“An empirical performance and programmability study has been performed by researchers.[SUP][45][/SUP] The authors claim that to achieve high performance Xeon Phi still needs help from programmers and that merely relying on compilers with traditional programming models is still far from reality.”
Still, it’s a dream, though. Phi is very, very interesting to me.