Preparing for GPU3: all nVidia GPU clients will soon need CUDA 2.2 or later to fold

By , the in Folding@Home Project - 2 Comments
The GPU client has had a great impact on the Folding@home project throughout its lifetime. In the latest publication produced by the project, (which can be found here or here in video form) it has been said that GPUs have enabled the project to study simulations many times longer than was previously possible, or indeed practical, and has made it possible to simulate protein folding for longer than one millisecond, for the first time using Molecular Dynamics.

The team is now preparing the deployment of the new GPU3 client. This new project will further enhance the contributions of GPUs to the project, allowing more precise simulations and analyses that can be performed twice as fast as previously possible. The new core should also be far more stable, as well as easier to incorporate new methods of analysis into. This is possible due to the use of the OpenMM library, which was created from the code at the heart of the current GPU cores, and improved upon significantly by the Simbios team.

The first phase of the deployment of the OpenMM GPU3 core (core 15) for nVidia GPUs requires the donator to perform two tasks:

1) For the core to function at all, your computer needs to support at least CUDA 2.2 and preferably the newer 2.3 version. Later versions are preferable if possible. To determine the version of CUDA you have on your machine, please consult the list of drivers and their included CUDA versions below:

  • CUDA 2.0: ForceWare 177.35 or higher (the minimum for the current GPU2 client)
  • CUDA 2.1: ForceWare 180.60 or higher
  • CUDA 2.2: ForceWare 185.85 or higher (minimum for the new GPU3 core [core 15])
  • CUDA 2.3: ForceWare 190.38 or higher (recommended for the new GPU3 core [core 15])
  • CUDA 3.0: ForceWare 195.36 or higher (probably the minimum version for the second phase of GPU3, which will use OpenCL)

2) A new client will be required in order to download GPU3 WUs. This new client will report the installed version of CUDA to the assignment servers, to ensure that your system is assigned only units which it can fold. However, to avoid finding yourself without work if GPU2/CUDA2.0 units are in short supply, be sure to upgrade your drivers to the latest available version. The Pande Group take the view that it is better to assign no work to a machine than risk erroneous results or too many EUEs due to old CUDA versions being installed.

Vijay has made this announcement now to give folders time to update their drivers before the newer drivers are required.

The Pande Group wishes to thank all of the folders who have contributed such a large amount of work, and whose results are yet to appear publicly.

Updated drivers can be downloaded from the CUDA Zone (English only) or by choosing your region here.

Adapted from: The Official Blog