NVIDIA has open sourced its CUDA compiler for the benefit of "academic researchers and software-tool vendors." To enable them "to more easily add GPU support for more programming languages and support CUDA applications on alternative processor architectures". The NVIDIA compiler is based on the popular LLVM
The NVIDIA CUDA is a GPGPU(General-Purpose GPU) solution that enables software to take advantage of a computer's graphics hardware for non-graphics related tasks. The kind of highly parallelized computation that graphics hardware excels at can also be beneficial in processing other kinds of data. Video encoding and decoding, mathematical / scientific calculations, and other tasks that can benefit from massive parellization could be run on the GPU to greately accelerate throughput.
Currently CUDA is restricted to NVIDIA hardware only, however with the this opening of the CUDA platform it could be used for other architectures, such as those by Intal and AMD. An alternative open standard OpenCL, by the creators of OpenGL is also available and can be used across platforms.
The NVIDIA compile source code is not available openly -- at least not right now -- but only for "qualified academic researchers and software tools developers," who need to register at NVIDIA's website in order to obtain access to the source code.