[Beowulf] getting a phi, dammit.

[Mark Hahn]

> At least with the Phi, within a single system, its completely
>cache-coherent... all 60 cores (240 hardware threads).

well, hardware-managed, so it's still time-slicing the core's resources.
(one FPU per core, for instance.)

> Am I stupid in thinking that they would also make very interesting web
> applications servers - get your database on a solid state array 
> Beside the Phi and crank up those web servers.

sure.  in HPC terms, webserving is usually just cache-friendly serial
farming, and scaling it is really just a matter of providing lots of 
modest-speed integer-based CPUs with decent amounts of memory bandwidth.

Phi only has 8GB ram, which for 60 real cores or 240 timesliced ones,
is not that much.  320GB/s theoretical is 5.3 GB/s/core - not all that
great either, but decent considering this is a many-slow-cores approach.
(for comparison, Calxeda puts 4x 1.4GHz arm cores on a ddr3 dimm - 
probably 2.7 GB/s/core, but a lot higher ram capacity/core.)

I don't know how well Phi manages IO, but its 16x PCIe v2 interface
has a theoretical peak of 8GB/s each way, which seems plenty as long
as the 8G ram onboard has a decent hit rate.

my guess is it would make a great webserver except for the 8G limit.