Nvidia released the latest CUDA 2.2 Beta, which contains a host of significant new features including the first hardware debugger for the GPU. The latest CUDA 2.2 Beta includes:

Hardware debugger for the GPU
Linux developers can now use a debugger on CUDA-enabled GPUs that offers both the familiar interface of the open-source GDB debugger and the ability to debug kernels as they execute on the GPU. The GPU-side debugger has all the features that developers expect from GDB, including the ability to have breakpoints, watch variables, inspect state, etc., as well as additional functions for CUDA-specific features.

Visual Profiler v2.2 for the GPU
The most common step in tuning application performance is profiling the application and then modifying the code. The CUDA Visual Profiler is a graphical tool that enables profiling C applications running on the GPU. This latest release of the CUDA Visual Profiler supports full measurement of memory bandwidth within a kernel, giving developers visibility into one of the most performance-critical areas of CUDA.

Full support for Microsoft Windows Server 2003/2008
Tesla C1060 and S1070 are now fully supported under both Microsoft Windows Server 2003 and 2008, offering developers and high-performance computing users more flexibility in their choice of operating system. CUDA 2.2 runs on Windows, MacOSX and major LINUX distributions.

Additional Features Coming with CUDA 2.2

- Improved OpenGL interop performance
- Texture from pitch linear memory
- Zero-copy support for direct access to system memory
- Pinned shared system memory allows compute kernels to share system memory
- Asynchronous memcopy on Vista

For a full list of new features in CUDA 2.2 and to get early access, see discussion on the CUDA forums:
http://forums.nvidia.com/index.php?showtopic=92580