17 Replies Latest reply on Nov 12, 2010 8:31 AM by RRick_77

    CUDA Cores Speculation and a Question

    zm15 Level 1

      I know most of us won't know this for some time, but I'm curious if at some point PPro with support more CUDA cores. From what i've read it appears that about 100 are supported now. Has anyone seen or heard differently?

       

      Now here's my question, is this something as 'simple' as unlocking more cores via an update to the code that can be passed down through the adobe updater? Or is this something that has to be overhauled into a next version?

       

      I don't know the slightest about coding, or what this type of work this involves, but I do understand the marketing side of it

        • 1. Re: CUDA Cores Speculation and a Question
          SteveHoeg Adobe Employee

          Premiere Pro's CUDA acceleration has always supported more than 100 CUDA cores.

          • 2. Re: CUDA Cores Speculation and a Question
            zm15 Level 1

            Thanks for the reply Steve.

             

            The reason I was thinking only 100 cores were used was because in most test's that i've seen there is no dramatic improvement between say a GT250 and a GTS450, even though the GTS450 has nearly 100 more cores than the GT240 - 96 vs 192.

             

            Where is the limitation or bottleneck with the MPE if you have a fast machine? ( i.e. i7 6 core, RAID 0's, etc..) Why do the better video cards not scale with better specs?

            • 3. Re: CUDA Cores Speculation and a Question
              Colin Brougham Level 6

              The reason I was thinking only 100 cores were used was because in most test's that i've seen there is no dramatic improvement between say a GT250 and a GTS450, even though the GTS450 has nearly 100 more cores than the GT240 - 96 vs 192.

               

              You do realize, of course, that the two (or is it three?) GPUs you've listed here are unsupported in terms of hardware accleration in Premiere Pro CS5? Unless you're using one of the officially supported and recommended cards, describing the lack of acceleration is without base and merit.

               

              The list of supported cards (at this time) includes:

               

              • Quadro 5000
              • Quadro 4000
              • Quadro FX 5800
              • Quadro FX 4800
              • Quadro FX 4800 Mac
              • Quadro FX 3800
              • GeForce GTX 470
              • GeForce GTX 285

               

              You're not going to get an official declaration on what's happening under the hood unless you're using one of these.

              • 4. Re: CUDA Cores Speculation and a Question
                zm15 Level 1

                Correct, I was using those two as a reference for test's that i've seen in the past, but the same holds true even between the gtx285 and gtx470, there doesn't seem to be any added benefit for the higher end gtx470. Take a look at the MPE Performance Chart at the top of the page here: http://ppbm5.com/Benchmark5.html

                 

                The average results are nearly identical for these two cards. I am taking them with a grain of salt I undertand, I'm just curious to know of the MPE can use all the CUDA cores on a card?

                • 5. Re: CUDA Cores Speculation and a Question
                  John T Smith Adobe Community Professional & MVP

                  Premiere using Cuda cores for Mpe is a combination of several things... Premiere code AND the nVidia driver AND the exact footage and/or effects being used

                   

                  I don't think you will ever get a "simple" answer to what is a complex interaction of at least those 3 factors (and, there may be other factors I have not mentioned 'cause I haven't thought of them)

                  • 6. Re: CUDA Cores Speculation and a Question
                    Colin Brougham Level 6
                    I was using those two as a reference for test's that i've seen in the past

                     

                    Got it. But as mentioned previously, I don't think it's worth considering unofficially "working" cards, because the software is obviously not being tuned to take advantage of them the same way that the officially supported cards are. If they were, they'd be on the list too

                     

                    the same holds true even between the gtx285 and gtx470, there doesn't seem to be any added benefit for the higher end gtx470.

                     

                    That's a more valid argument, as far as I'm concerned. However, I think one needs to consider that this is a pretty new technology, and that it hasn't been fully optimized yet. The Mercury Playback Engine--not even consider the hardware acclerated component of it--has only been around for less than a full product cycle. This isn't based on any sort of knowledge in particular--just on how the world works--but I think it stands to reason that it's going to get "better." At that point, you'll see the advantage of a newer/bigger/faster GPU. Such is the price of being on the bleeding edge...

                     

                    The average results are nearly identical for these two cards. I am taking them with a grain of salt I undertand, I'm just curious to know of the MPE can use all the CUDA cores on a card?

                     

                    An engineer would have to answer that (if they want to, or can ), but I'd repeat my reply above: it's a work-in-progress.

                    • 7. Re: CUDA Cores Speculation and a Question
                      zm15 Level 1

                      Thanks for the insight Colin, I'll add my +1 to all your comments.

                       

                      Any engineers care to join on this discussion and shed a little light??

                      • 8. Re: CUDA Cores Speculation and a Question
                        Harm Millaard Level 7

                        I'm not an engineer, but can show you some figures rendering the following:

                         

                        PPBM5 benchmark, MPEG2-DVD timeline, overlayed with 6 additional tracks of H.264-BR sequences, each starting at different timecodes, each with different opacity and rendering that takes around 70 seconds on my system with a GTX-480.

                         

                        Memory used maxes out at 1363 MB, GPU load maxes out at 66% (with the next highest score being 44% and on average around 25-30%), memory controller load maxes out at 33% (average around 15%).

                         

                        These figures show me that the video card is not taxed very much, only memory is used to capacity. If these conclusions are correct and more than 100 CUDA cores can be used by design, the bottleneck will be the maximum VRAM and the regular system components.

                        • 9. Re: CUDA Cores Speculation and a Question
                          gtacyclist Level 1

                          I use EVGA's Precision app which can monitor GPU temps, activity (load), among other things. On my 470 I only see the GPU taxed around 10% to 30% with an average of perhaps around 20% when I am rendering (I usually have 3 main video layers each with around 2 to 3 MPE accelerated effects - Flip Vertical, Flip Horizontal, Drop Shadow). This to me means "only" around 25% of the cores are being used which translates to 100-120 or so out of the 470's 448 cores. I suspect adding more accelerated effects would just slow down rendering instead of taxing the GPU more.

                          • 10. Re: CUDA Cores Speculation and a Question
                            Harm Millaard Level 7

                            Slight correction to your statement:

                             

                            This to me means "only" around 25% of the cores are being used which translates to 100-120 or so out of the 470's 448 cores.


                            75% of the time CUDA cores are idle, awaiting data from the CPU, system memory or VRAM or the bus (think disks) to hand the data over.

                            • 11. Re: CUDA Cores Speculation and a Question
                              zm15 Level 1

                              Those are really interesting finds gtacyclist and Harm, thanks for sharing those.

                               

                              So it appears that that only a percentage of the card's memory is being used. Is there a way to tell via your programs how many 'cores' are being used?

                               

                              I suppose one could make the argument that although the VRAM isn't being used, the shaders and all the cores are. Valid hypothesis??

                               

                              I'm not sure how things get processed, but I would imagine someone close to the science that is the MPE could answer.

                              • 12. Re: CUDA Cores Speculation and a Question
                                gtacyclist Level 1

                                EVGA's Precision tool does show the amount of VRAM being used. VRAM usage hovers around 600-650 MB when rendering the work area and around 500-550 MB when doing realtime previews with the output resolution on my second monitor (a 19" widescreen with 1440x900 resolution) set to full (instead of 1/2 or 1/4). This probably varies with the project type of course... my main video files consist of 2 x 720p AVHCD footage at around 3 hours each plus a real time screen cap footage using Microsoft's Expression Encoder 4 API.

                                 

                                Conversely to the VRAM usage, the cores do seem to get used more when doing real time previews... constant 20-22 percent GPU utilization than when rendering the work area which tasks the GPU anywhere from 5 to 25% (CPU is tasked more on rendering).

                                 

                                From what I can see, VRAM and the CUDA cores are not being utilized to the max by PPro CS5 on my project. I know this for a fact because when I play a game, GPU utilization (and temps) jump to a much higher level than what I see when using PPro CS5.

                                 

                                I am not complaining by any means... but had I known before what I know now I would probably have settled on a cheaper, more efficient, cooler running, less noisy GPU... most likely a GTX 460 instead of a 470. The advantage of a 470 of course is there is no need to apply the hack everytime Adobe releases a patch/update.

                                • 13. Re: CUDA Cores Speculation and a Question
                                  zm15 Level 1

                                  gtacyclist wrote:

                                   

                                  ...

                                   

                                  I am not complaining by any means... but had I known before what I know now I would probably have settled on a cheaper, more efficient, cooler running, less noisy GPU... most likely a GTX 460 instead of a 470. The advantage of a 470 of course is there is no need to apply the hack everytime Adobe releases a patch/update.

                                   

                                  Would you have gone for a card like the GT240 since it's been shown that all the VRAM doesn't get used and the higher end cards don't get used to their potential? That is, if you didn't do much gaming.

                                  • 14. Re: CUDA Cores Speculation and a Question
                                    gtacyclist Level 1

                                    I probably would have gone with the GT 460 since it is the closest performance wise to the card it replaced which is a ATI Radeon 5850. I do play games from time to time and IMHO, the GTX 460 / 5850 (now superceded by newer models) represented the best bang for the buck.

                                     

                                    Having said that, the 470 has dropped in price significantly the past week or so and can be had for 259 not including mail in rebate offers on some models (ASUS). As I said the advantage to the 470 is that it is "officially" sanctioned by Adobe so there is no need to perform the hack to enable MPE acceleration.

                                    • 15. Re: CUDA Cores Speculation and a Question
                                      RRick_77

                                      bump

                                       

                                      Any engineers for input?

                                      • 16. Re: CUDA Cores Speculation and a Question
                                        Todd_Kopriva Level 8

                                        RRick 77 wrote:

                                        > Any engineers for input?

                                         

                                        What is your question?

                                        • 17. Re: CUDA Cores Speculation and a Question
                                          RRick_77 Level 1

                                          Todd_Kopriva wrote:

                                           

                                          RRick 77 wrote:

                                          > Any engineers for input?

                                           

                                          What is your question?

                                          I'm curious if perhaps there is any input from Adobe as to why there doesn't seem to be much disparity between the entry level MPE cards and the higher end cards, as they have very different price points for very little gain on the higher end cards.