Announcement

Collapse
No announcement yet.

NVIDIA Shows Off "Kayla" Running On Ubuntu

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #11
    Originally posted by EpochDC View Post
    OpenCL is the industry's standard, CUDA is a proprietary Nvidia technology.
    Though it has a very large market penetration, it's not the standard...
    Industry standard != open specification. Windows, DirectX and Exchange are all proprietary and in the same time are industry standards.
    Look at this link: http://blogs.amd.com/fusion/2011/01/...stry-standard/
    Microsoft offers ways to integrate in embedded their solutions: http://www.microsoft.com/windowsembe...solutions.aspx

    I put MS here, but are many other industry standards that are standards but not open. Java for a long time was not open, and even OpenJDK is the open specification, certainly the Oracle's JDK is the Industry standard and most distros and OSes do install the Oracle's Java and not OpenJDK (you're free though to install the OpenJDK).

    OpenCL though is a better standard to bet into future: if AMD (or eh, Intel) will bring tomorrow a much better compute GPU, you can switch to it even you bought today NVidia. With CUDA you will be always stuck. Also theoretically at least (it depends on the final device OS support) OpenCL allows to split the computing tasks through all cores and GPUs, so if you have a very fast CPU and a mediocre GPU you can "accelerate" the computations on your CPU, when again if you're using CUDA, your CPU will simply wait for your GPU to finish all computations.

    Comment


    • #12
      Originally posted by ciplogic View Post
      Industry standard != open specification. Windows, DirectX and Exchange are all proprietary and in the same time are industry standards.
      Look at this link: http://blogs.amd.com/fusion/2011/01/...stry-standard/
      Microsoft offers ways to integrate in embedded their solutions: http://www.microsoft.com/windowsembe...solutions.aspx

      I put MS here, but are many other industry standards that are standards but not open. Java for a long time was not open, and even OpenJDK is the open specification, certainly the Oracle's JDK is the Industry standard and most distros and OSes do install the Oracle's Java and not OpenJDK (you're free though to install the OpenJDK).

      OpenCL though is a better standard to bet into future: if AMD (or eh, Intel) will bring tomorrow a much better compute GPU, you can switch to it even you bought today NVidia. With CUDA you will be always stuck. Also theoretically at least (it depends on the final device OS support) OpenCL allows to split the computing tasks through all cores and GPUs, so if you have a very fast CPU and a mediocre GPU you can "accelerate" the computations on your CPU, when again if you're using CUDA, your CPU will simply wait for your GPU to finish all computations.
      You are confusing things
      Industry Standard is: Generally accepted requirements followed by the members of an industry.

      Comment


      • #13
        You're right, I was too quick to claim that OpenCL is the industry's standard.

        Though neither of the two seem to be the standard, Opencl has a broader support from vendors. (Which is natural given the proprietary nature of CUDA)
        And CUDA has a larger market share in both consumer and server/HPC markets...

        IMHO, opencl will prevail , Nvidia's (hardware) market share is not large enough to force a propietary, closed standard upon the industry.

        Losing this 'standard war' would not be disastrous for Nvidia, since their GPGPU's support opencl as well.
        And CUDA can co-exist next to opencl for a long while (as it does now): since it's easier to implement for software developers, it will stay the nr. one choice for many of them.

        Comment


        • #14
          The standards war is already lost for Nvidia...

          Not only most hardware supports OpenCL, but since as the time passes developers will use OpenCL more, they will get to know it better. CUDA's marketshare was mostly based on its headstart. Nvidia took GPGPU seriously far earlier than other hardware vendors. But CUDA is doomed and anyone saying otherwise is wrong...

          Comment


          • #15
            So this is basically what AMD is doing with their new Tablet processors, but ARM based. Maybe the GPU part here is more powerful, I didn't check. The power consumption of the CPU part, however, should be comparable.

            Comment


            • #16
              We can't know about performeance and power consumption, but feature-wise (64bit) only parker seems to be a true competitor for kabini.

              Edit: nvidia seems to be ahead on the driver department though, by already supporting android and gnu/linux, whereas amd questionably does the later and don't support the former, so they are behind software-wise.
              Last edited by Figueiredo; 20 March 2013, 10:29 AM.

              Comment


              • #17
                Nobody writes CUDA code for consumer applications, that's true.
                But in number crunching industries, people really do use CUDA much more than openCL. That's because of commercial support and better optimization for Nvidia cards, and because Tesla cards are the only server dedicated GPGPU with ECC memory on the market (hence, only Nvidia cards are used on GPGPU grids anyway).
                Maybe it's different in public and semi-public research (more openCL because open/standard), I have no particular experience on that.

                Comment


                • #18
                  Tegra 4 was late by several months, yet they think they can do Tegra 5 for 2014Q1, and it's a bigger jump in many ways. Won't be holding my breath for that to release on time.

                  Comment


                  • #19
                    Originally posted by curaga View Post
                    Tegra 4 was late by several months, yet they think they can do Tegra 5 for 2014Q1, and it's a bigger jump in many ways. Won't be holding my breath for that to release on time.
                    Dealing with TSMC appears to be a huge burden.

                    Comment


                    • #20
                      Originally posted by johnc View Post
                      Dealing with TSMC appears to be a huge burden.
                      Then they should build their own semi plant. If TSMC is unreasonable they should switch vendors or build their own, but complaining about it isn't getting them any sympathy.

                      Comment

                      Working...
                      X