NVIDIA announced at GTC Asia in Bejing that they are opening up the CUDA platform by releasing their compiler surce-code. The new NVIDIA CUDA compiler is based upon LLVM.
NVIDIA says they are open-sourcing their CUDA compiler for "enabling [researchers and software vendors] to more easily add GPU support for more programming languages and support CUDA applications on alternative processor architectures."
The LLVM-based CUDA compiler is part of the CUDA Toolkit 4.1 release. Along with releasing the source-code, they are providing documentation on the CUDA IR (Intermediate Representation) format.
CUDA can now come to other processing architectures and even other non-NVIDIA GPUs.
This move isn't entirely surprising though based upon the growing presence of OpenCL and it becoming widely adopted across many CPU and GPU architectures.
While the CUDA compiler is now open-source, getting the actual source-code is a bit difficult at the moment. NVIDIA is making "qualified academic researchers and software tools developers" fill out a form, but when doing so you don't even have immediate access to the code. With that said, it's not known at the moment under which license NVIDIA is releasing the CUDA compiler.
NVIDIA announced this open-source move in a press release and comes just days after AMD open-sourced its IL code generator for OpenCL, which also utilizes LLVM.