There is a very interesting article on G80/G92 hardware architecture just published by NVIDIA guys (hardware architects) in the March/April edition of IEEE Micro:

NVIDIA Tesla: A Unified Graphics and Computing Architecture

Erik Lindholm, NVIDIA
John Nickolls, NVIDIA
Stuart Oberman, NVIDIA
John Montrym, NVIDIA

The full article can be downloaded here: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4523358

Most of the informations can be found or deduced from the CUDA manual but there is also few details about hardware model in this article that was not given before as directly by NVIDIA. Especially, there is mention of Fragments grouped into 2×2 to be scheduled into Warps, a thing we deduced with our G80 hardware analysis (http://www.icare3d.org/GPU/CN08).There is also the firstpicture of the G80 die publicly avalable (hum ok, not very clear and in grayscale) !