The next version of the CUDA SDK, version 3.0, is beginning to show. Available only for the moment among some developers (who received e-mail containing information to download), this SDK improves support for OpenCL and supports cards Fermi.
Improvements in the technology itself include:

• Management of double precision
• Interoperability between OpenCL and OpenGL for better display performance
• Recovery of Compute Capability via cl_nv_device_attribute_query
• Possibility to control the compiler optimizations through cl_nv_compiler_options
• Support for images OpenCL filtering better and faster
• Supports 32-bit atomic operations
• Support Addressable Bytes Stores
• Support the revision 1.0.48 of Khronos specifications OpenCL
• Support OpenCL headers of the Khronos 1/11/2009
In the SDK and ToolKit, we find the following news:
• Interoperability between the driver and the CUDA runtime buffer, which allows applications using the API to use CUDA drivers also use the Runtime libraries CUDA C
• A new version of the CUDA C Runtime (CUDART) for debugging mode emulation
• Support C + + (class inheritance and templates) to facilitate the life of the developer
• New unified API for Direct3D and OpenGL for:
- Interoperability of OpenGL textures
- Interoperability of Direct3D11
• Support for debugging hardware cuda-gdb for those using the CUDA driver API
• New tool for checking the available memory in cuda-gdb and as a tool in its own right
• The versions of the CUDA Toolkit libraries are now set, allowing applications to operate a particular version or to support multiple versions
• The core C / C + + CUDA are now compiled in the format ELF
For future cards Fermi announced but not yet available include:
• The native 64-bit
• The Copy Engine Multiply
• Error handling ECC
• The Concurrent Execution Kernel
• Debugging hardware via the Fermi-cuda gdb

All these innovations should be publicly available within weeks, but active developers have therefore already have access to these innovations. Recall that CUDA technology is currently available in version 2.3, Linux, Windows and Mac OS X, and it allows the use of tools like C for CUDA, CUDA for Fortran, etc.. Through the processors feed GeForce (since the 8th generation), which NVIDIA has recently changed the name to core CUDA.