The software developer toolkit that NVIDIA first launched years ago has gone through several transformations, the latest of which was revealed less than 24 hours ago.
The CUDA Toolkit, as the company decided to name its creation, made quite the leap in feature support and performance.
From the 4.2 version, released back in April (2012), it jumped all the way to version 5.0. A sharp contrast with the advancement from 4.0 (May 2011) to 4.1 (January 2012) and then 4.2.
The software development toolkit provides C and C++ application creators with a few new possibilities.
For one, all device functions can be separately compiled and linked using NVCC. This permits the creation of closed-source device function libraries, and even lets those libraries initiate user-defined device callback functions.
The linker support is a BETA in this release, but customer feedback will enable NVIDIA to iron out whatever bugs may be left.
A second feature of CUDA Toolkit 5.0 is the new command-line profiler, nvproof. It lets optimization efforts be properly focused by providing summary information about where applications spend the most time.
A third asset of the release is the CUDA Dynamic Parallelism, which lets GPU-run global and device functions launch kernels using the “<<<<>>>>” syntax and call CUDA Runtime API routines directly. Granted, this ability existed before, but only from host functions.
The fourth and final new feature is Nsight Eclipse Edition for Linux and Mac OS, an integrated development environment user interface that lets programmers develop, debug and optimize CUDA code.
All in all, CUDA Toolkit 5.0 opens the door for new and improved programs capable of harnessing the parallel computing capabilities of graphics processing units.
Developers can download the appropriate version of the software from one of the links below.