While NVIDIA should soon be releasing a new Linux graphics driver beta, for those of you interested in NVIDIA's Compute Unified Device Architecture (CUDA) rather than -- or as a complement to -- OpenCL, there is a new tool-kit release. CUDA 3.2 is now available this week. CUDA 3.2 brings a number of new features to the NVIDIA GPGPU table.
The November 2010 release of CUDA Tool-kit 3.2 includes 50 to 300% CUBLAS performance improvements for Fermi GPUs (GeForce 400 / 500 series), CUFFT performance tuning on Fermi, new GPU libraries for accelerating new routines, the H.264 encoding and decoding libraries are now bundled with the CUDA tool-kit, multi-GPU debugging support, NVCC support for the Intel C Compiler 11.1 on 64-bit Linux, support for memory management using the malloc() and free() functions within CUDA C compute kernels, and various other changes. The NVIDIA System Management Interface (nvidia-smi command) is now also capable of reporting how busy the GPU is as a percentage, which is a feature we like for the Phoronix Test Suite and benchmarking to be able to monitor GPU load. We've already been able to do this on the AMD side on Radeon hardware via the OverDrive extension and somewhat with their open-source drivers via monitoring the fence count.
The NVIDIA CUDA 3.2 release for x86/x86_64 Linux can be downloaded at NVIDIA.com
. This is meant for developers and includes the latest NVIDIA 260.xx Linux beta graphics driver.