3 Replies Latest reply on May 3, 2013 1:07 AM by himanshu.gautam

    OpenCL implementation roadmap ?

    phollensen

      I just had the pleasure of running some tests on a 7970 over at hgpu.org.

      Amazing how much simpler you can make things and still get 2.5 TFlops when you have 8MB of registers!

      Cudos, AMD.

       

      From what I gather though, there are some current limitations that aren't  documented anywhere.

      For instance according to http://devgurus.amd.com/message/1287981#1287981 ,

      asynchronous DMA transfers work on PCIe 2.0 but not yet on PCIe 3.0.

      Nuggets of information like this are spread thoughout this forum, making it very difficult to know what to expect.

       

      So I'm hoping to get an idea of what issues are on your radar,

      and whether they will have to wait on other projects you consider more pressing, such as the Bolt Library.

       

      Another big one is peer-to-peer communication.

      From what I gather, the 7990 you are releasing implements Crossfire over a PLX PCIe 3 bridge.

      This offers a very low latency path for gpu-gpu communication that, I think, would be boon for many of us.

      Are there any intentions to make that available for gpgpu in light of this new hardware release?

       

      In general, it would be very helpful if I could follow an issue and be notified when its resolved,

      and also see the number of followers it has so I know if there's broad interest in a feature or if I'm in a small niche.