[frugalfirbolg]

Can anyone comment on using Intel's software suite (Parallel Studio, Cluster Studio) for parallel programming? A workflow comparison to CUDA would be particularly interesting. From what I see in the programming examples, the Intel notations are used to identify code that is run on the Phi instead of the host and the compiler automagic handles the details, versus what I've seen of CUDA where you have to explicitly copy to the card's memory and manage the blocks.

Some points I found in [ref 1]: - each Xeon Phi card is running Linux and has it's own IP - you can SSH into the individual Xeon Phi just like hopping into a headless box - you can write apps that run entirely on the Xeon Phi

I don't see a physical ethernet jack on the card, so even though it has a network stack, it looks like it has to go through the host to access a NAS or other resources. I guess if network I/O becomes your bottleneck you're better off with separate boxes still.

I also wonder what prevents you from slapping this card into a system that doesn't have a Xeon processor. Is it a technical limitation such as having only developed PCI express drivers for that chipset so far, or marketing? I guess buying one and watching the driver install would reveal some answers if it's not stated in the literature.

[ref 1] http://software.intel.com/sites/default/files/article/330164...