Intel Xeon Phi

[RichAlgeni]

http://www.tomshardware.com/reviews/xeo ... ,3342.html

When something like this becomes available, do they introduce it with new assembler instructions?

Would it be something on the order of calling a library or dll to utilize the power of this? If such a library or dll were available?

[DarkDragon]

In essence, Xeon Phi takes 60 (at least in the announced 5110P SKU) x86 cores with big 512-bit vector units, runs them in excess of 1 GHz, and yields more than 1 teraFLOPS of double-precision performance on a dual-slot PCI Express card with a custom Linux distribution

You might be able to code in PureBasic for those extension cards. That's the main advantage compared to CUDA and OpenCL.

Instead, ISVs can optimize for Phi using C, C++, and Fortran, with specific additions [...]

I think it only uses some kind of thread building blocks/openmp addon for those languages.

[idle]

By the sounds of it no you wouldn't need any specific api's, you'd just need a compiler that
has the target optimizations for the coprocessor. From what I understood they've been using GCC
and running it on Redhat and Susi.

[RichAlgeni]

Thanks folks. Don't envision needing it anytime soon, but I was just wondering about it.

[DarkDragon]

By the sounds of it no you wouldn't need any specific api's, you'd just need a compiler that
has the target optimizations for the coprocessor. From what I understood they've been using GCC
and running it on Redhat and Susi.

But you may need the api to get the x86 program onto the extension card, as it would be somehow crazy to schedule everything smoothly through the bus. And there's a separate OS installed on those cards - they would have implemented it on a lower level if they wanted to schedule automatically between this device and the mainboard processor.

[idle]

Yes you would undoubtedly need an API to load a program and interact with it for the setting and getting
They had just said something about the same parallel code developed for an i3 would run without change on the Xeon Phi
which would be a lot easier to test and debug, then you'd just need to add the code to do the marshalling.
Though while it's all well and good to have x86 based code, what tools do they have in mind to assist in developing
effective parallel code? It'd probably be easier to use opencl without a specific api for development on both x68 and the Xeon.