GPU3: no support for ATI in the 1st version

By , the in Folding@Home Project - 3 Comments
GPU3: no support for ATI in the 1st version
Vijay Pande has just published a progress review of the new GPU3 core, which is currently under development. As a reminder, this core will be based on OpenMM, a library developed by Stanford to accelerate and improve the simulations, while making code maintenance and updating much easier. OpenMM supports OpenCL beta in version 1.0 (scheduled for late January 2010), but it should be noted that OpenCL is still very young, and it will not be present in the first version of the GPU3 core. This first core should still use CUDA on nVidia cards, and CAL/Brook on ATI cards.

As CAL/Brook is no longer maintained by ATI, the core will initially be available only for nVidia GPUs. Once OpenCL implementations have matured (both ATI and nVidia), developers will be able to complete and optimize the OpenCL code, then those portions of code will finally leave test quality.

If you've been waiting on OpenCL as some kind of messiah (particularly in terms of support for ATI GPUs), then for your religious fervour you will be punished, at least with the first version, but once the environment has matured, developers can move forward and once again do ATI's work for them. Moreover, when the support for multi-core machines by OpenCL has also matured, the development team will consider porting GPU3 code to the SMP2 code (that is to say, develop a GPU core based on SMP threads). If performance is good, it may be possible to envisage a return to a common codebase, and have most calculations performed on a single Folding@Home platform.

Update - 14/01, 22hr UTC :
Vijay has made some corrections and clarifications in his post on the OpenCL implementations. It should first be noted that the developers are having problems with implementations of both ATI and nVidia, which probably explains why the first version of the GPU3 core so closely resembles the current GPU2 version, and its use of CUDA. Vijay does not yet know if the bugs originate within the Folding@Home code, or from the implementations themselves, but the Stanford team is working closely with nVidia, ATI, and even Apple, to ensure full code compatibility with each implementation.

The ATI and nVidia implementations are fully functional; only the Folding@home OpenCL code is not currently fully working.

Adapted from: Some more details on the GPU3 core, regarding OpenCL