CUDA is supported by 2 architectures : the 2 years-old GeForce 8xxx/9xxx/GT/GTS architecture and the new GTX architecture.
There’s finally few differences between the new GTX architecture and the old one, such as doubling the register number, and the coalesced memory access algorithm, the double (64bits floating point number) support, and few instructions.
Even on the first architecture, older GeForce 8800 (CUDA 1.0 architecture) may lack some features of the actual ones (or Geforce 9800/GTS250 as they were renamed 
), such as atomic operations. Cruel!
I will focus on what is known as CUDA 1.1 devices, such as formet GeForce 8800, 8600, 8500, any 9xxx or GT/GTS. They share the same general architecture and metrics, mainly differenciated by number of SP (Scalar Processors), frequencies, memory bandwidth (bus size & frequency) and naturally memory size.
I have one GeForce 8600M GTS, with 32 SP (Scalar Processors) on my laptop, and a GeForce 8800 GTS, with 96 SP on my home desktop, that is approximately 4X faster, due to increase in frequency as well as having more cores.
I plan to go for a GeForce GTX 260 (CUDA 1.3 device) if I am lucky on CraigsList or find a good sale, in the future, while staying compatible and optimizing for the CUDA 1.1 devices.
Tags: CUDA 1.1, CUDA 1.3, CUDA Architecture, GeForce 8600M GTS, GeForce 8800 GTS