Nvidia released CUDA Toolkit 2.2 to all developers today. As I mentioned in April when the beta shipped to registered developers, the toolkit includes a visual GPU profiler that now works on Vista (see figure below), a GPU debugger that now works on 64-bit Linux as well as 32-bit Linux, and a handful of new APIs introduced in response to developer feedback.
[ Keep up with app dev issues and trends with InfoWorld's Fatal Exception and Strategic Developer blogs. ]
The version that shipped today also supports Windows 7, has improved OpenGL interoperability, saves bandwidth by taking textures from pitch linear memory, avoids memory copy overhead by letting CUDA functions read and write directly from pinned system memory, extends asynchronous memory copying to Windows Vista, and allows an application to claim exclusive use of a GPU.
I didn't have any trouble downloading or installing the new device driver, toolkit, and SDK. I'm not in a position to verify the claimed performance improvements over CUDA 2.1; what I can say is that the new version feels very solid on my hardware.
You can download the CUDA 2.2 device driver, toolkit, and SDK from the Nvidia site.
