7 Replies Latest reply on Feb 24, 2012 1:45 AM by craft_coder

    OpenCL Radeon support

    vaspa

      Hello,

      I have been working with OpenCL and GPUs from just 1-2 months so my first question is going to be kind of dum.

       

      I have now a computer with installed AMD APP SDK v2.5 and an ATI Radeon X1650 GPU but it seams that I cannot create a platform with my GPU.

       

      I saw in the requirements page http://developer.amd.com/sdks/AMDAPPSDK/pages/DriverCompatibility.aspx that there is not GPU from the Radeon X1000 series.

       

      Is this right?

      Does this mean that with no implementation and no drivers will I be able to use my GPU to execute kernels?

       

      Thank you in advance and best regards.

        • Re: OpenCL Radeon support
          MicahVillmow

          You must have a GPU based on the R7XX architecture or later for use with OpenCL. The Radeon X1000 series is so old it is not even supported by CAL or Brook+.

            • Re: OpenCL Radeon support
              craft_coder

              On a related note, I was waiting to buy a Pitcairn-based card (HD 78xx) and got the bright idea that maybe I should pick up a cheap Cypress-based unit (HD 58xx) for the near-term, under the theory that any optimizations I do for it should easily apply to the GCN architecture.

               

              Is that true?  If so, it seems I should definitely target the more challenging architecture, since VLIW-5 and VLIW-4 cards will continue to comprise much of the installed base, for a few more years.

               

              The main reason I see not to do that would be any anticipated lack of support for OpenCL 1.2 on the Evergreens.  Can you comment on that possibility, at all?

                • Re: OpenCL Radeon support
                  MicahVillmow

                  OpenCL 1.2 will work on Evergreen and later hardware.

                   

                  In most cases, code optimized for Evergreen hardware will scale well to GCN architecture. If this is not the case, we would love to see an example so we can figure out what is causing the scaling to not occur.

                    • Re: OpenCL Radeon support
                      craft_coder

                      Thanks for the prompt reply!

                       

                      So, how am I to compare between the two (i.e. if I'm to tell you when an optimization hurts GCN)?  I mean, when I buy a GCN card, should I expect that I can install it along side a HD 58xx, in the same machine?  Or would I need a second machine?  I'm running Linux (Ubuntu 11.04 64-bit on a Phenom II / 890FX, with PCIe lanes aplenty! :-)

                       

                      I (mostly) wouldn't be interested in running on both devices, concurrently.  Just targeting one or the other, for benchmarking & testing purposes.

                       

                      BTW, I'm currently running on a HD 5450 that I bought while awaiting rollout of the GCN cards.  It's useful for basic development/testing purposes, but that's about it.

                        • Re: OpenCL Radeon support
                          notyou

                          craft_coder wrote:

                           

                          Thanks for the prompt reply!

                           

                          So, how am I to compare between the two (i.e. if I'm to tell you when an optimization hurts GCN)?  I mean, when I buy a GCN card, should I expect that I can install it along side a HD 58xx, in the same machine?  Or would I need a second machine?  I'm running Linux (Ubuntu 11.04 64-bit on a Phenom II / 890FX, with PCIe lanes aplenty! :-)

                           

                          I (mostly) wouldn't be interested in running on both devices, concurrently.  Just targeting one or the other, for benchmarking & testing purposes.

                           

                          BTW, I'm currently running on a HD 5450 that I bought while awaiting rollout of the GCN cards.  It's useful for basic development/testing purposes, but that's about it.

                          As I understand it, GCN moves a lot of the work that had to be done by the compiler for VLIW to hardware. Since more of the scheduling can by done dynamically you should have increased performance in most if not all cases (I'll leave it to the AMD guys to further this since they know the most about the architecture).

                           

                          You should have no issues installing multiple GPUs from different architectures in the same machine. The only thing you need to worry about is selecting the correct device in your program and, if you were going for something more exotic (but you're not), load balancing since the two architectures are different.

                          • Re: OpenCL Radeon support
                            MicahVillmow

                            I currently have, or have had in the recent past, a 7970, 5770 and a 4350 in one machine and a 5970, 6970 and a 5550 in my other machine. As I need to move cards around quite often for testing it changes about weekly, but there should not be any problems with running multiple generations of cards in the same machine.

                    • Re: OpenCL Radeon support
                      craft_coder

                      Thanks, notyou & Micah.

                       

                      I opted for the HD 6850.  There are some fantastic deals out there, and it's a plenty powerful card if you don't mind optimizing for VLIW-5 and the lack of double-precision.  Though I'd prefer a few more SIMD cores (it has 12), it should be wide enough for my purposes.  The 256-bit memory interface was an absolute requirement, for me.  All the floating point horsepower in the world wouldn't be much good if you can't feed it.

                       

                      I wanted a HD 5850 (18 SIMD cores), but very few new units are available & I'd rather not waste my time debugging a flaky used board that was overclocked and/or under-cooled by some G4M3R kid.

                       

                      I hope to post more about my projects, in coming months.  I'm already far enough along to be anxious for some real hardware.