Archive notice: This article was originally published on November 19, 2007. Links and embedded videos are preserved as part of the historical record.
ith the new CUDA release 1.1, NVIDIA’s standard display drivers now also include a special CUDA driver. This eliminates the installation of specific software, allowing developers to deploy CUDA-capable applications easily in various markets. In addition, version 1.1 of the software development kit includes new source codes with which users can perform CUDA calculations on multiple GPUs. CUDA 1.1 for Windows is now available as a free download at http://www.nvidia.de/cuda, CUDA 1.1 for Linux is available from November 19.
In addition, NVIDIA is presenting a new CUDA debugger at SuperComputing 2007 that is based on a standard GDB interface. The debugger allows break points to be set and variables to be set/checked directly via the graphics processor, and all of this in the C language. NVIDIA’s debugger also supports a standard graphical DDD interface. Third-party debuggers can be layered seamlessly over CUDA.
CUDA is already used by 20 universities worldwide for parallel programming, including the University of Illinois in Urbana-Champaign: “Our scaling efficiency across multiple GPUs is almost perfect – four GPUs give us exactly four times the performance for our calculations”, says John Stone, Senior Research Programmer at the University of Illinois. “A single GPU offers 290 GFlops; with four GPUs, the performance can be expanded to 1,156 TFlops, a scaling efficiency of 99.7 percent.”
Other universities relying on NVIDIA’s CUDA: University of California, Davis; Stanford University; University of Pennsylvania; University of California, Berkeley; University North Carolina at Chapel Hill; Purdue University; John Hopkins University; University of California, San Diego; India Institute of Technology, Delhi.
Further links:
- Nvidia/Cuda