NVIDIA CUDA Toolkit 3.2 available
NVIDIA, maker of graphics processors (GPUs) and chip-sets, has released a new version of its developer package for writing CUDA (Compute Unified Device Architecture) programs. Version 3.2 of the CUDA Toolkit introduces new mathematical libraries and contains performance revisions and better cluster management. The GPU debugging and performance analysis tools have also been improved.
According to the manufacturer, the previously included libraries such as "Basic Linear Algebra Subroutines," and "Fast Fourier Transform" were already up to 20 times faster when compared to Intel's Math Kernel Library (MKL). The new libraries, such as "curand" for the generation of random numbers and "CUSPARSE" for matrix functions, should be much faster than MKL. Finally, the toolkit supports the development and / or encryption of H.624 video, and integrates with the Tesla Compute Cluster (TCC).
CUDA is a programming model and a development environment that allows programmers to easily make use of the processing power of NVIDIA GPUs. CUDA based extensions to MATLAB and Photoshop plug-ins are also available. External projects support the connection to Python. NET, Fortran or IDL. CUDA can be used to customise programs to achieve a performance boost with almost all of the current NVIDIA product lines – GeForce, Quadro and Tesla.
- CUDA support in OpenCV announced at GTC, a report from The H.
- Yellow Dog Linux for CUDA updated, a report from The H.