I worked on my laptop (GeForce 8600M GTS that is a CUDA 1.1 Device) and I began to port it to my home desktop that have a GeForce 8800 GTS.
And I begin to have some stupid errors about non-supported atomic operations…
I take a look to check how my GeForce 8800 is recognized, and I discovered that is it a GeForce 8800GTS of first generation, that is a CUDA 1.0 DEVICE, and CUDA 1.0 DEVICES doesn’t include atomic operations, complicating the inter-thread communication (kinda nearly impossible on some case!).
This is why I *HATE* marketing names, because a GeForce 8800 is not necessarily compatible with a GeForce 8800, and with the software developped for it, as it is between GTX 260 and GTX 260M.
I thought I have been cautious, but now, I will have to change it, probably for a CUDA 1.3 DEVICE, such a desktop GeForce GTX 260, that will enable me to develop targeting this architecture, or wait for firsts G300 cards that may appear in september, and will probably implement a better set of instruction, specific optimizations and probably more register or cache.
