12 Replies Latest reply on Apr 16, 2011 10:44 AM by nou

    Is the memory limit being raised eventually?

    galmok

      Can we expect the memory allocation block limit to be raised beyond 128MB (on a 1GB 5870) and beyond 512MB (total)? I am asking this because we are doing some calculations that require much memory and well, 128MB isn't going to do it.

      Just for comparison, I checked the GTX470 we have (1.28GB total memory) and we can allocate and use (verified) 1.1GB in one continous block of memory even though the GPU is the only display adapter in the pc (being used by windows 7 64-bit at the same time).

      In that light, 128MB is a bit small...

      I am aware of the environment variables, but they shouldn't really exist in the first place. Also, using them leads to unpredictable results.

        • Is the memory limit being raised eventually?
          MicahVillmow
          Yes it will be raised, expect it in the upcoming release.
          • Is the memory limit being raised eventually?
            MicahVillmow
            Sorry, we don't give out ETA's on when the release will be coming.
            • Is the memory limit being raised eventually?
              ryta1203

               

              Originally posted by: galmok Can we expect the memory allocation block limit to be raised beyond 128MB (on a 1GB 5870) and beyond 512MB (total)? I am asking this because we are doing some calculations that require much memory and well, 128MB isn't going to do it.

              Just for comparison, I checked the GTX470 we have (1.28GB total memory) and we can allocate and use (verified) 1.1GB in one continous block of memory even though the GPU is the only display adapter in the pc (being used by windows 7 64-bit at the same time).

              In that light, 128MB is a bit small...

              I am aware of the environment variables, but they shouldn't really exist in the first place. Also, using them leads to unpredictable results.

               

              Yep, AMD really dropped the ball on this one...

              ...I'm not sure how they can even talk about doing GPU computing with this limitation... letting Nvidia slide that much farther away in the GPGPU race.

                • Is the memory limit being raised eventually?
                  Meteorhead

                  The estimated time to arrival always being referred to as "soon" reminds me of the summer kenue tours, where the day's finish is always "right after the next turn".

                  It's not complaint, but there is always this great secrecy about new SDK release dates, and the reason for us asking about it so many times is: we just can't wait to see it!

                  • Is the memory limit being raised eventually?
                    galmok

                     

                    Yep, AMD really dropped the ball on this one...

                     

                    ...I'm not sure how they can even talk about doing GPU computing with this limitation... letting Nvidia slide that much farther away in the GPGPU race.

                     

                    To be fair, nVidia themselfes are delaying their official OpenCL 1.1 release. It is currently only available to registered developers and only works up to the 400 seriers cards (and is 10 month old now). They haven't voiced any information about when it will be released and as such, it could look that AMD will be the only one with a proper GPGPU-supported OpenCL 1.1. Officially, nVidia only has OpenCL 1.0...

                      • Is the memory limit being raised eventually?
                        ryta1203

                         

                        Originally posted by: galmok


                         

                        Yep, AMD really dropped the ball on this one...

                         

                        ...I'm not sure how they can even talk about doing GPU computing with this limitation... letting Nvidia slide that much farther away in the GPGPU race.

                         

                        To be fair, nVidia themselfes are delaying their official OpenCL 1.1 release. It is currently only available to registered developers and only works up to the 400 seriers cards (and is 10 month old now). They haven't voiced any information about when it will be released and as such, it could look that AMD will be the only one with a proper GPGPU-supported OpenCL 1.1. Officially, nVidia only has OpenCL 1.0...

                        AMD could have OpenCL x.x++ (whatever) and it wouldn't matter since you can't operate on anything larger than 128MB, I'm just saying it's a HUGE limitation that AMD has been talking about fixing for awhile now but hasn't.

                          • Is the memory limit being raised eventually?
                            nou

                            digging in libamdocl.so again and this what i got.

                            export GPU_MAX_HEAP_SIZE=100
                            export GPU_MAX_ALLOC_PERCENT=100

                            and clinfo report 1GB global memory and 1GB max allocated memory.

                              • Is the memory limit being raised eventually?
                                galmok

                                 

                                Originally posted by: nou digging in libamdocl.so again and this what i got.

                                 

                                export GPU_MAX_HEAP_SIZE=100 export GPU_MAX_ALLOC_PERCENT=100

                                 

                                and clinfo report 1GB global memory and 1GB max allocated memory.

                                 

                                 

                                But you can't use it... :-/

                                  • Is the memory limit being raised eventually?
                                    rollyng

                                     

                                    Originally posted by: galmok
                                    Originally posted by: nou digging in libamdocl.so again and this what i got.

                                     

                                     

                                     

                                    export GPU_MAX_HEAP_SIZE=100 export GPU_MAX_ALLOC_PERCENT=100

                                     

                                     

                                     

                                    and clinfo report 1GB global memory and 1GB max allocated memory.

                                     

                                     

                                     

                                     

                                     

                                     

                                    But you can't use it... :-/

                                     

                                    Hi, just curious, have you proved that it does not work despite clinfo shows all memory?

                                    Would you like to share your tes code? Thanks

                                      • Is the memory limit being raised eventually?
                                        nou

                                        i succesefully increased memory up to 700MB througth GPU_MAX_HEAP_SIZE and GPU_MAX_ALLOCATION_SIZE. but only when i allocated and used only one buffer. when i tryed two buffers with size 2x300MB second one was allocated in main memory. it works but it was slow.

                                        i created simple write kernel to this buffers. when i allocated one 700MB buffers it write ~90GB/s. when i used two buffers first have ~90GB/s and second one only ~5GB/s so IMHO it was allocated in host memory. my observation is that buffers are not allocated on device if you fill original 50% memory limit.