Dally must have got round to reading the CS books from the 1990s
This is my favourite from 1992 :
http://docs.google.com/viewer?url=http%3A%2F%2Fwww.ulb.tu-da...
His is spot on though, but we need to get the vector unit as standard on the MB, not some add-on card. Larabee talks good graphics talk but the real gift should just be the vector FPU made available to the OS.
SSEx processes 128bits in one operation which can represent 2 x 64bit doubles or 4 x 32bit floats or 8 x dwords or 16 bytes, which I thought was me reading it wrong, they call it SIMD and let you have 2 64bit data points which I suppose is technically true :) I was expecting something spectacular like 128.