Archive notice: This article was originally published on March 11, 2014. Links and embedded videos are preserved as part of the historical record.
Nvidia is now offering version 6 of its CUDA architecture for parallel computing as a free download. Here is an overview of the version highlights:
Unified Memory
The most important innovation allows CUDA applications to access CPU and GPU without data having to be copied back and forth manually. This is intended to save a good deal of time and make it easier for pogrammers to integrate GPU acceleration into applications.
Drop-in Libraries
These are intended to enable simple acceleration of applications by up to a factor of 8x. The new drop-in libraries automatically accelerate BLAS and FFTW calculations by replacing existing CPU-only BLAS or FFTW libraries with new GPU-accelerated equivalents.
Multi-GPU Scaling
Redesigned BLAS and FFT GPU libraries automatically scale performance across up to eight GPUs in a single node. This gives one node up to nine teraflops of double-precision computing power and supports larger workloads (up to 512 GB).
Further information and download options are available here.
Image rights (C): NVIDIA
