20 Replies Latest reply on Mar 7, 2013 9:38 AM by Dave Knarr

    GPU Question


      I bought a GTX 550 ti for $100 - but then I saw that the 650 ti is only about $150 or so, and I was wondering if the 650 ti is worth me taking back the 550 to get the 650 - basically is the 650 MUCH better, or just slightly better?  Hope that makes sense, and thanks.

        • 1. Re: GPU Question
          Jeff Bellune Level 6

          [moved to hardware forum]

          • 2. Re: GPU Question
            Harm Millaard Level 7

            The 650 Ti is not as good as the 550 Ti, so keep it.


            The memory bus is smaller (128 versus 192 bits)  and the memory bandwidth is lower (86.4 GB/s versus 98.4) so there is no benefit whatsoever.

            • 3. Re: GPU Question
              Bill Gehrke Most Valuable Participant

              I definitely would do it.  The 550 Ti has only a 192 CUDA cores but the 650 Ti has 768 CUDA cores.  For only $50 more it is a good deal.  It does somewhat depend on the rest of your system.  Cores are king in my testing.

              • 4. Re: GPU Question
                Harm Millaard Level 7

                Carefull Bill!


                Otherwise you may need to explain why a 670 and a 580 score equal, or a 560 and a 660 score equal, both despite huge differences in cores, but equal in memory bandwidth. I've said it before:


                Video cards performance

                With the number of observations available now (1190+) we see a clear tendency that more CUDA cores do help performance, at least with the older architecture. The usual performance increase of hardware acceleration over software is around 12 x, with the cards with fewer CUDA cores lagging behind and the ones with more cores pulling ahead. The increase in performance over software shows that indeed the 9800 GT and GTX 260 are clearly lagging when compared to the GTX 480/580/680. The newer architecture shows its benefits.

                On the other hand, with the introduction of Kepler cards with around three times the number of CUDA cores over the Fermi cards, people expected triple the performance and they were disappointed to be shown wrong in their expectations.

                With all the results we have from our PPBM5 benchmark it sure looks like memory bandwidth may be the decisive factor to impact performance. It is no longer just number of CUDA cores.

                • 5. Re: GPU Question
                  Bill Gehrke Most Valuable Participant

                  Unfortunately, I do not have enough money to have all the cards, but supposing I test my: 

                  1. GTX 560 Ti with 448 CUDA cores 320-bit wide 152 MB/seconds transfer rate.
                  2. GTX 660 with 960 CUDA cores with only a 192-bit wide 144MB/second memory transfer rate

                  Will the results be significant to see what is really happening?

                  • 6. Re: GPU Question
                    Harm Millaard Level 7

                    My guess is they will show about equal.

                    • 7. Re: GPU Question
                      Bill Gehrke Most Valuable Participant

                      Harm sorry to disappoint you but they are not equal, but as usual you are correct in that the GTX 560 Ti beats a GTX 660!  In our Direct Export MPEG2-DVD test the GTX 560 Ti in ten successive scores encoded the timeline in 40 seconds average versus the GTX 660 that averaged 45 seconds for that same process!  You learn something new every day..  Since these two boards are so quite similar there may be more to it than just memory bandwidth


                      This is just not my week  I have been wrong, twice now but, with our benchmarking capability I have achieved my real end goal of being able to prove to myself and the world an opinion presented on these forums is or is not true. In this case my opinion even with all my many, many hours of testing that I have done did not lead me to the correct conclusion.


                      I can see my next prioity will be more GPU testing.

                      • 8. Re: GPU Question
                        Harm Millaard Level 7



                        Thanks for reporting back. These differences are bigger than I expected, even with the measurement errors Windows give, but over ten runs a difference of more than 10% is bigger than I would have expected which was around 6% based on memory bandwidth alone. Nevertheless, it proves my point that it is not only about CUDA cores, but more about memory bandwidth, and that is the main reason that I think the Titan card will be around 30 - 40% faster than a 580/680 card, provided the rest of the system is up to it.

                        • 9. Re: GPU Question
                          RjL190365 Level 4



                          Thanks for the comparison. As for your findings, note that the "GTX 560 Ti" result is only for the 448-core version (which is a GTX 570 with one of its 15 texturing units disabled). The standard GTX 560 Ti is a different GPU - one with only 384 CUDA cores and a 256-bit wide memory bus with only a 128 MB/second transfer rate.

                          • 10. Re: GPU Question
                            Bill Gehrke Most Valuable Participant

                            Randall, thanks for pointing that out that my GTX 560 Ti 448 is a slightly crippled GTX 570 (maybe they had some yield problems???).


                            Here are a couple of other numbers that I quickly came up with:. 

                            • GTX 680 (stock clock) the ten tries provided an interesting pattern.  35, 34, 35, 34, 35, 34, 35, 34, 35, 34 so I guess that one is 34.5
                            • GT 640 (384 CUDA cores, 128-bit bus width, 28.5 GB/s 2GB DDR3 RAM) was slow so only 3 runs:  148, 147, 148 seconds  (because it is past my bedtime)


                            --more later, it will also be interesting to see the results of 1 GB boards because this test has been using 1.2GB of video RAM

                            • 11. Re: GPU Question
                              RjL190365 Level 4

                              Thanks again, Bill. It seems as though with the Kepler architecture, memory bandwidth is king.


                              One more thing: The GTX 560 Ti 448 wasn't the only "crippled" GTX 570 ever manufactured. There was another, OEM-only version of the GTX 560 Ti (that came with some Dell PCs) that was even further crippled than the GTX 560 Ti 448: This one had another three texturing units disabled (on top of the one unit that had been disabled in the GTX 560 Ti 448), resulting in that OEM GTX 560 Ti having only 352 CUDA cores. All other specs (clock speed, memory bus width, memory transfer rate) were identical for all three derivatives of the GTX 570.

                              • 12. Re: GPU Question
                                medeamajic Level 2

                                The memory bandwidth is important but the GTX 600 series and the GT 600 series support 4 monitors. I opted for a GTX 650 Ti. It works just fine for my needs.  I tested it out with some of my older monitors (4 monitors total) to see if it would work. It worked OK but my old monitors are 4:3. I will buy at least one more 16:9 monitor soon.

                                • 13. Re: GPU Question
                                  Jim_Simon Level 9

                                  it sure looks like memory bandwidth may be the decisive factor to impact performance. It is no longer just number of CUDA cores.


                                  I won't say memory bandwidth has no effect, but I'm not sure it's the 'decisive' factor.  The 570 has more bandwidth than the 660, but fewer cores.  The 660 wins (albeit barely).


                                  It's likely that cores, bus and bandwidth play interconnected roles.

                                  • 14. Re: GPU Question
                                    Harm Millaard Level 7

                                    Jim, have a look at Video Card Performance and look at the chart. That says it all. And the 660 loses by around 12.5% from the 560 Ti, let alone the 570.

                                    • 15. Re: GPU Question
                                      RjL190365 Level 4

                                      Thanks, Harm, for that chart (based on Bill's own testing). The GTX 660 may have 960 CUDA cores - but the shader units in the Kepler GPUs run at exactly the same clock speed as the GPU core (unlike Fermi which runs its shaders at double the GPU core clock). This effectively cuts the GTX 660 performance to the same level as an otherwise equally clocked Fermi GPU with "only" 480 CUDA cores. So, with the GTX 660's GPU performance effectively equal to that of an older Fermi 500-series GPU with half the CUDA cores, it's easy to see how the GTX 560 Ti 448 still outperforms the GTX 660. If on the other hand the "standard" GTX 560 Ti is tested against the GTX 660, the standard 560 Ti would lose to the GTX 660 by a rather significant margin: In addition to the standard 560 Ti having only 384 CUDA cores (versus the 480 effective Fermi-level CUDA cores in the GTX 660), the standard GTX 560 Ti has a memory bandwidth of only 128 GB/s (versus 144.2 GB/s in the GTX 660).


                                      And the GTX 480, despite more CUDA cores and higher memory bandwidth, still loses to the GTX 560 Ti 448 because the GTX 480 runs at a lower clock speed than the GTX 560 Ti 448.


                                      As for the GTX 650 Ti, it is not sufficient enough of an improvement over the GTX 550 Ti to justify the exchange. The GTX 650 Ti has more than sufficient CUDA cores (384 effective, relative to Fermi GPUs; 768 actual) to overcome the memory bandwidth deficit versus the GTX 550 Ti. However, the lower memory bandwidth still limits its performance to no higher overall than a plain, non-Ti GTX 560 (with its 336 CUDA cores and the same 128 GB/s memory bandwidth as the standard GTX 560 Ti). Simply put, the GTX 650 Ti could have been quite a bit faster than it actually is.

                                      • 16. Re: GPU Question
                                        Bill Gehrke Most Valuable Participant

                                        Now that we have an entry on a Titan that essentially shows no noticeable improvement over the GTX 680 all theories seem to go out the window!  I believe that we have reached a point of diminishing returns.

                                        • 17. Re: GPU Question
                                          Harm Millaard Level 7

                                          Another warning Bill ,


                                          When the 680 came out, expectations were high due to the tripled number of cores, but the very first submissions did not prove that to be true. Now we have one Titan result (which has not been submitted BTW) and it seems a bit early to say it makes no discernable difference. I still feel the performance increase over a GTX 680 is in the region of 30 - 40%, but that is a gut feeling, nothing to substantiate it. But BFTB wise, it does not make sense to get such a costly card at this moment.

                                          • 19. Re: GPU Question
                                            Jim_Simon Level 9

                                            Jim, have a look at Video Card Performance


                                            Curious.  I was looking at the following.



                                            • 20. Re: GPU Question
                                              Dave Knarr Level 1

                                              Hi Jim,


                                              In my article you are refering to, the CPU's are both a bit dated.  I will be adding some new benchmarking results with newer CPU's in the next week or two.  I am still running tests.



                                              What I am noticing is on the newer CPU's the 570 and the 660 are very close in results and the 570 is actually a few seconds faster with some tests.



                                              Dave Knarr