NVIDIA has released version 2.2 of the CUDA toolkit and SDK. The new version of CUDA profiles C coded applications to use the computers GPU in addition to the CPU. This can enhance the computer's overall processing speed. The update also includes the ability to zero-copy, or to read and write directly from pinned system memory rather than send it through the graphics memory system. Using the technique requires either a GeForce 9300/9400 (or 9400M) integrated graphics core or else a GeForce GT 200 series board. CUDA 2.2 also lets multiple GPUs access the same system memory information. This feature, Pinned Shared System, results in less overall memory being used as multiple processors access the same bits. The update aims to deliver improved performance of OpenGL applications running on Quadro GPUs by splitting OpenGL graphics functions on to different GPUs. In the Windows version, support for Windows 7 is also new; Mac OS X support remains and sees all the new features.
Using most features requires a GeForce 8 series graphics chipset or later.
Apple hopes to eventually promote a transition from vendor-specific technology like CUDA to OpenCL, its self-created but universal standard for general-purpose computing on GPUs and multi-core CPUs. More here.