Announcement

Collapse
No announcement yet.

NVIDIA Releases CUDA 6.5 As A Huge Update

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

  • NVIDIA Releases CUDA 6.5 As A Huge Update

    Phoronix: NVIDIA Releases CUDA 6.5 As A Huge Update

    NVIDIA put out version 6.5 of their Compute Unified Device Architecture today and it is a big step ahead, including better development tooling for CUDA Fortran...

    http://www.phoronix.com/vr.php?view=MTc2ODk

  • #2
    C++11 support

    NVIDIA hasn't publicized it, but most importantly, CUDA 6.5 includes C++11 support

    Comment


    • #3
      They seem to have dropped OpenCL support and want us to use their proprietary CUDA. Their GTX cards despite being expensive enough are optimized for games/gamers only. AMD has much better OpenCL performance but worse Linux drivers, Blender not working. Also in youtube flash and fullscreen the vid lags. AM/NV have both their flaws (NVIDIA because they are artificially restricting their cards and AMD because they don't care about Linux).

      Comment


      • #4
        Originally posted by opensource View Post
        They seem to have dropped OpenCL support and want us to use their proprietary CUDA. Their GTX cards despite being expensive enough are optimized for games/gamers only. AMD has much better OpenCL performance but worse Linux drivers, Blender not working. Also in youtube flash and fullscreen the vid lags. AM/NV have both their flaws (NVIDIA because they are artificially restricting their cards and AMD because they don't care about Linux).
        Agree, for OpenCL readon dirver is not ready, and Catalyst have its shortcommings.
        (However, they did improved few releases before, and they are targeting HSA now, which may mean better OpenCL at last!)

        But I have no problems with vids nor with flash (in Chrome!), using r600 driver on my old 5730M.

        Comment


        • #5
          Originally posted by opensource View Post
          They seem to have dropped OpenCL support and want us to use their proprietary CUDA. Their GTX cards despite being expensive enough are optimized for games/gamers only. AMD has much better OpenCL performance but worse Linux drivers, Blender not working. Also in youtube flash and fullscreen the vid lags. AM/NV have both their flaws (NVIDIA because they are artificially restricting their cards and AMD because they don't care about Linux).
          AMD was having a problem even getting OpenCL support to the point where it could run even basic features in Blender in Windows. It may have improved, but about a year ago, it was more or less a lost cause because subroutine support was so inefficient that the driver couldn't get compute kernels compiled once they hit a certain complexity (running out of RAM especially). It's partially an architecture issue as I believe nvidia has better hardware features to assist this, but a lot of it is that really fixing it for AMD gpus requires a compiler to act in more smart ways than it would have to for nvidia hardware, at least to achieve those aims without obliterating performance.

          I suspect that when linux+amd+opencl support really gets there, it will be coinciding with LLVM getting better at some of these compilation tricks. Probably just in time for some new technology that won't be supported for 5 years

          Comment


          • #6
            Originally posted by opensource View Post
            They seem to have dropped OpenCL support and want us to use their proprietary CUDA. Their GTX cards despite being expensive enough are optimized for games/gamers only. AMD has much better OpenCL performance but worse Linux drivers, Blender not working. Also in youtube flash and fullscreen the vid lags. AM/NV have both their flaws (NVIDIA because they are artificially restricting their cards and AMD because they don't care about Linux).
            It's not that AMD doesn't care about Linux. Their Windows drivers suck balls too. My desktop has a 560 Ti. Then I bought a couple of 280x for mining and just for funzies I swapped one of them with the nvidia to see how much better it would perform in a game. I didn't do benchmarking per se, just wanted to see if there's an difference noticeable by the naked eye. After finishing my mini-experiment I reverted back to nvidia, that's when the display issues started. Prior to this event I never had an ATI/AMD card before and I was surprized a bit to see that uninstalling catalyst is not a trivial task in Windows. The driver shortcomings have been dragging on for a long time now (inheritance from ATI). AMD's focus right now is on the HSA research which, if anything, is an indication that they might shift to the open-source kernel-space drivers and user-space closed source catalyst sooner rather than later to save themselves from spreading out resources too wide.

            Comment


            • #7
              http://arstechnica.com/apple/2013/06...new-mac-pro/2/

              Apple may be trying to not-so-delicately nudge everyone to move their code from CUDA to OpenCL, but Iíve seen a first-hand failure of AMDís OpenCL support with V-Ray RT for Maya. Chaos Group built V-Ray RT on OpenCL, but after extensive work trying to get the GPU variant of its RT render engine running on AMD hardware and an effort by yours truly to light a fire under Apple and AMD, Chaos Group gave up and ported it to CUDA instead. So V-Ray RTís GPU mode only works with OpenCL and CUDAóon Nvidia hardware.
              http://www.luxrender.net/forum/viewt...p?f=34&t=11009

              This is the result with Intel OpenCL CPU device:
              4.5 seconds is a quite reasonable time.

              This is the result with AMD OpenCL CPU device:

              I don't know how much takes to compile the kernel because I gave up after 15 minutes.

              I mean ... seriously ... C compilers for CPUs are a 40 years old technology.
              http://www.luxrender.net/forum/viewt...11009&start=10

              AMD GPU compiler is far far worse than the CPU one. My surprise here was in the finding that even CPU compiler has problems.
              AMD OpenCL is pure garbage, you cannot rely on AMD if you actually do real work.

              Comment


              • #8
                Originally posted by opensource View Post
                They seem to have dropped OpenCL support and want us to use their proprietary CUDA. Their GTX cards despite being expensive enough are optimized for games/gamers only. AMD has much better OpenCL performance but worse Linux drivers, Blender not working. Also in youtube flash and fullscreen the vid lags. AM/NV have both their flaws (NVIDIA because they are artificially restricting their cards and AMD because they don't care about Linux).
                AMD have better performance with simple kernels, but when you try something like the OpenCL Myocyte test, that is larger than a mining kernel, the performance drop makes AMD 5x to 8x times slower than nvidia on both linux and windows.

                AMD seem to have dropped OpenCL support some years ago and want us to use HSA that is for APUs only.

                Comment


                • #9
                  Originally posted by opensource View Post
                  They seem to have dropped OpenCL support and want us to use their proprietary CUDA. Their GTX cards despite being expensive enough are optimized for games/gamers only..
                  The GeForce series were always meant to be a gaming-only card ever since it was first launched.

                  Want 3D cards for professional work, Quadro. And Quadro only.

                  Comment


                  • #10
                    Ok.

                    The GeForce series were always meant to be a gaming-only card ever since it was first launched.

                    Want 3D cards for professional work, Quadro. And Quadro only.
                    I don't like that NVIDIA artificially restricts their consumer cards (at least much more than AMD) so I'd probably try an AMD card because of this first.

                    AMD seem to have dropped OpenCL support some years ago and want us to use HSA that is for APUs only.
                    HSA? but aren't APUs slower than dedicated big GPU cards?

                    But I have no problems with vids nor with flash (in Chrome!), using r600 driver on my old 5730M.
                    I used their proprietary since it's usually faster. I used FF, chrome has its own flash plyer AFAIK.

                    Then I bought a couple of 280x for mining
                    No comparison against ASICs (which are optimized for doing one thing only (hashing of one specifig lengh) whereas GPUs are a bit like FPGAs (programmable, to work with different lengh)).

                    Comment


                    • #11
                      Originally posted by opensource View Post
                      Ok.
                      HSA? but aren't APUs slower than dedicated big GPU cards?
                      Yes, but AMD OPENCL have huge performance drops with large Opencl kernels (If it's too large it will fail to compile, but it'll work with Intel and nvidia Opencl).

                      it's a known bug reported ~3 years ago that has not been fixed yet even when they said it's a high priority.

                      Comment

                      Working...
                      X