beelsebob wrote:I would be very surprised if they were 4 cores. The E5 Xeon series has very few 4 cores, all of which are poor value. The 6 core models obviously can't make 8 core machines, so I'd assume you're looking at things like the E5-2687W, which are 8 cores per CPU. Though then I'd be kinda surprised that they didn't stick two of them in a board. All in all, sounds a little odd.
It is (or at least was when I was last involved in building one) usual in HPC clusters to use several different kinds of node in a single solution: you shouldn't take the total and divide it by the node count and expect the result to be the specs of every node.
For example, you might use pairs of six-core processors in "compute" nodes (that do the maths), these will have super-fast network connections to each other; the priority is calculations per second per box for these. At the other end of the spectrum you can use pairs of 2-core processors in your "i/o" nodes (which aggregate data and store it to disk), these are doing very little processing, basically just co-ordinating a DMA network input from compute nodes and a DMA disk output to disk. In Caterham's case they could usefully designate a third class of "render" node to visualise the data; these would likely be light on processing power and incorporate good GPUs for hardware-accelerated video compression if nothing else.
The currently-new generation of HPCs are much-more GPU-based: both Nvidia and Intel make special boards for this purpose which are essentially a top of the range graphics chip with no video outputs. You can get, on one board, an amount of computational power that would have got you to the top of the worldwide supercomputer charts in ~2001
