i’m just wondering, i know that there are alot of threads about it, but i’m more curious on the cycles side, and the compatibility side, would it even run well?
because, a gefore is though cheaper, it has less usefulness, meanwhile, the xeon is loaded with usefulness cuz it’s like another system in your system, so you can give them stuff to compile.
so how does both of them stack up in terms of raw performance, my system is xeon CPU, so it’s compatible
Get the GeForce. Eventhough the Xeon Phi 31S1P is currently available for around 100$ - 200$ inside the US (awkwardly, no export to EU private persons; no EU offers - this offer seems to be US only), you should still get a GeForce. While the Xeon Phi is perfect for scientific calculations, which depend on double precision performance and ECC memory, you will be perfectly fine with a usual GeForce card. Single precision performance is needed for Blender.
Plus, the CUDA code is well tested by several thousands of blender users, while there doesn’t even exist a Xeon Phi version. Honestly, as long as you are not a developer and not familiar with programming at all, just keep away from the Xeon Phi. It is not intended for individuals, but for super computer clusters, universities, scientific applications and so on. It is only made for a very specific workload and can’t be used in a wide variety of applications. You will not benefit from buying a Xeon Phi; just get a decent GeForce and enjoy the other benefits, like widespread software support (Photoshop, Premiere, PhysX…) and gaming performance.
Performance wise, at best optimization and highly parallel, non-memory bound calculations, it looks like the Xeon Phi is about 1.5x - 2.5x times faster than a 2x 8C/16T Xeon system (sorry, I don’t have source atm). Else, it is inferior or at best on par with the Xeon system.
“An empirical performance and programmability study has been performed by researchers.[SUP][47][/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.” http://en.wikipedia.org/wiki/Xeon_Phi
basically you would need to write a custom compiler or rewrite the code. if you have the money to buy one maybe having custom software written for you wont seem too expensive. ask intel how much it would cost then decide if the total cost is worth it for you. i would go with the custom compiler over custom code. as fast as blender gets updated your custom code would be behind while a custom compiler migh let you stay up to date but its probably going to be more expensive.
The new version of the Intel compiler does some hefty legwork making things run faster on the Phi, but unfortunately I have no Phi to test on. I can whip up an ICC built version of Blender if someone happens to have one sitting around though.
But on the whole, yes, there would have to be some very specific programming done in order to get the most out of the Phi platform.
I have a Xeon Phi in my Workstation (Xeon Phi 31S1P). Let’s get in contact ;)!
oh?? so how does it perform?
@1832vin: I can’t tell for now, because I don’t have access to icc at the moment. I’ve received the Xeon yesterday, I didn’t have the time to do any benchmarks so far. Also, due to Blender beeing a huge project (depending on alot of other libraries and programs), I don’t think a simple recompile will work out-of-the-box.
There should be 2 options: either compile the Cycles OpenCL kernel for Xeon Phi or compile Cycles C++ directly. I’m new to MIC (primarily bought the Xeon Phi for educational purposes); so really can’t tell now, what it’s capable of.
let’s get in touch when you know!!
thankyou!