0 Replies Latest reply on Oct 6, 2011 9:52 AM by Meteorhead

    OpenCL linux issues

    Meteorhead
      someone tell me why?!?!

      Someone pls tell me two things:

      1: Why does it take nearly a year to make the drivers on linux be rid of XServer? XServer sucks soo bad, I cannot find the words. It is as instable as hell, and very easy to c-rash with badly conditioned code (not to mention it's true unneccessity).

      2: Why cannot OpenCL runtime ship with the drivers? When updating Windows drivers, the Catalyst install manager always notifies of updates, and the OCL runtime does get updated more often, than the SDKs themselves. Why must linux users wait 6 months with every fix after a messup??

      I'm only asking these two, because I'm getting real tired, of trying to see whether maybe multi-gpu has been fixed in the actual Catalyst drivers, and crashing the computer in 3 levels below in another building. I have asked this in a seperate topic, whether we should expect this to be fixed in an upcoming Catalyst or APP release, but nobody answered.

      It is so damn frustrating to know that with a single card installed (such as on my notebook), SDK 2.5 runtime brought over +50% increase in performance under OCL, and there is no way to get this working on the development machine with 3*HD5970 inside. Not to mention that even with reverting back to an older working pair of runtimes and drivers, programs crash that used to work.

      I would really much like to build a rackmount, dense GPU number cruncher, but honestly, I do not dare spend the money, because multi-gpu support is unreliable. With PCI-E 3.0 coming about, it wouldn't be so far off from reality to have 32 GPUs attached to a single node to achieve damned fast communication and low latency among many GPUs, but even if one overcomes BIOS restrictions, I feel that AMD simply lacks software support and reliability.

      This is not an insult to the developers, rather than a feedback to higher ups, that A LOT MORE human and financial resources should be allocated to this region. I know the hardware rocks, I can develop aplications that utilize them to their fullest, but I cannot recommend it to anyone with credits, because it is unusable for commercial use.

      I am doing research, where demand isn't as high, but if I were in the private sector, my **** would be seriously kicked if I advised someone to buy even a smaller cluster, and it has major bugs such as SDK 2.5, and must wait 6 months for a fix...

      At least release some 2.5.2 version SDK that fixes the multi-gpu issue...