nVidia is not well known in the business-computing world. The company is best known for its 3D graphics acceleration cards, and it is hoping to expand the appeal of those cards by offering developers tools that allow them to offload non-graphical computation tasks onto those cards. Yesterday, nVidia released version 3.0 of its CUDA parallel computing tool sets.
Chief among the updates are enhanced support for C++, a move to ELF binaries, and better tools for Linux users. Previous versions of CUDA did support C++, but now that support has been expanded to include inheritance for templates and classes.
Previous to version 3.0, CUDA applications were split out from standard applications at compile time. Developers simply mark the sections of code destined to run on the graphics processor, and the CUDA compiler wrapper splits off those sections, compiles them, then reinserts them into the fully compiled binary so that when run, the CPU and GPU would each get their own chunks of code automatically.
For version 3.0, those CUDA binaries are being converted from a proprietary format into the UNIX standard ELF format.
William Ramey, senior product manager for GPU computing at nVidia, said that CUDA is being used in corporations to add extra horsepower to desktops. Specifically, he said that developers are using CUDA to push nightly computation batches onto desktop machines.
CUDA also supports cards specifically designed for 1U and 2U enclosures. These nVidia graphics processors are embedded into cards specifically targeted at CUDA use, despite their similarity in architecture to commercially available graphics cards.
Ramey also said that before the end of the month, a CUDA-focused version of Microsoft Visual Studio will be made available to developers. This edition will include handlers for making CUDA development easier. The suite will be called Nsight, and developers will need to own a copy of Microsoft Visual Studio in order to use the Nsight add-on.