Announcement

Collapse
No announcement yet.

NVIDIA Launches The GeForce GTX TITAN X, Linux Tests Coming

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

  • #11
    Originally posted by efikkan View Post
    Well it's not a consumer card. Titan is intended for semi-pro users, who do compute, rendering and development. Previous Titan cards offered high DP performance and lots of memory, but unfortunately this version only offers more memory.

    Titan X have basically the same bandwidth as 780Ti with 3 GB. Even though games running at high resolutions with AA and high details easily could consume lots of memory, they will still be bottlenecked by the memory bandwidth. No game can use close to 12 GB in a single frame and maintain a good frame rate (do the math). Some games which uses "megatextures" will thrive with loads of memory, but unfortunately there are like one game or so. Things will start getting interesting next year with Pascal using HBM memory (and AMD later this year).

    As a developer myself, I would strongly consider buying one of these Titans. EVGA got a Superclocked version with 12% higher performance for only $30 extra.
    I agree there is a bandwidth limit internally to the card. However, this is a much faster connection than from main memory. Having that 12GB filled up will increase the speed of the card in these situations, compared to that information being in main memory. I am not saying the card will process it any faster, I am saying the card can access it faster, which will benefit the application, though not as much as if they had higher bandwidth on the card itself. I am also saying it is coming whether the card can handle it or not.

    Comment


    • #12
      Originally posted by dragorth View Post
      I agree there is a bandwidth limit internally to the card. However, this is a much faster connection than from main memory. Having that 12GB filled up will increase the speed of the card in these situations, compared to that information being in main memory. I am not saying the card will process it any faster, I am saying the card can access it faster, which will benefit the application, though not as much as if they had higher bandwidth on the card itself. I am also saying it is coming whether the card can handle it or not.
      True, however far too few games scale that well with memory capacity unfortunately. Even if you play like BF 4 and use close to 4 GB memory, a typical frame would probably use ~1.5 GB. The remaining 2.5 GB will be unused objects and parts of the landscape, and you will use different parts of this as you walk around the landscape. In theory, a game using only 1.5 GB per frame could still use all of the 12 GB to have a larger texture cache and keep the terrain in a higher detail level. For such use cases, the bandwidth would not be a big problem. Unfortunately, very few games are designed this way, and it's the chicken and the egg problem again. The only game I'm aware of i Rage, but I don't think it scales to anywhere near 12 GB.

      Comment


      • #13
        This is not a gamer card so even if what you say is true it does not apply.

        Titans are made for 3D compute in professionnal domains. A friend of mine use them to calculate special effects in advertisment or visual effects in tv programs for example. With a gamer card he would need time to compute then watch, with a few Titans he has a good enough fps to check the job in realtime.

        No gamers card can give him this performance because high levels / new effects are available and the amount of memory is usefull for scenes with a lot of elements when you want a correct result visually.

        Comment


        • #14
          Originally posted by Passso View Post
          This is not a gamer card so even if what you say is true it does not apply.

          Titans are made for 3D compute in professionnal domains. A friend of mine use them to calculate special effects in advertisment or visual effects in tv programs for example. With a gamer card he would need time to compute then watch, with a few Titans he has a good enough fps to check the job in realtime.

          No gamers card can give him this performance because high levels / new effects are available and the amount of memory is usefull for scenes with a lot of elements when you want a correct result visually.
          Except that this Titan, the Titan X, isn't. Those 3D professionals need the DP performance of the earlier Titans, and this one doesn't have it. This one is equivalent to the gamer cards in that respect. Those Titans can handle multiple billions of Double Precision, i.e. 64 bit Floating Point numbers, while this card can only handle about 770 million. In fact, this card will be 1/3 the speed of the previous Titan Black in terms of Double Precision. Whereas it's Single Precision numbers go from 5 Billion to 8 Billion operations per sec.

          So, this card IS a gaming card, compared to even the last year model.

          Comment


          • #15
            The Titan X is Maxwell based. It has an improvement of 20% in bandwidth usage efficiency.
            And the Titan X has more BW available than the first Titan (336 vs 288 GB/s), which happens to be the same than the 780 Ti and the Titan Black, but with that increase in efficiency.

            Comment


            • #16
              Originally posted by Filiprino View Post
              The Titan X is Maxwell based. It has an improvement of 20% in bandwidth usage efficiency.
              And the Titan X has more BW available than the first Titan (336 vs 288 GB/s), which happens to be the same than the 780 Ti and the Titan Black, but with that increase in efficiency.
              I am not sure whom you are replying to, but I will point out that better bandwidth does not equal better performance in all situations. In this case, without those DP compute units, that bandwidth is effectively useless for some subset of use cases, such as the aforementioned 3D Rendering.

              Comment


              • #17
                Originally posted by dragorth View Post
                I am not sure whom you are replying to, but I will point out that better bandwidth does not equal better performance in all situations. In this case, without those DP compute units, that bandwidth is effectively useless for some subset of use cases, such as the aforementioned 3D Rendering.
                DP is not needed for 3D rendering. DP is needed for science computations like physics simulation. Games are being 3D rendered in real time, and they are SP.
                2D video effects and filters, even if they are derived from 3D computations but without physics going underneath, run with SP.

                And you could still run physics simulations with SP for 3D rendering. The algorithms being used tolerate that. In fact I've been working with a physics simulator for 3D effects and floats (not doubles) were being used. Only in a small fraction of execution time it used doubles in order to not lose precision, but in my opinion that is stupid because the linear solver runs with float values.

                Comment


                • #18
                  Originally posted by Filiprino View Post
                  DP is not needed for 3D rendering. DP is needed for science computations like physics simulation. Games are being 3D rendered in real time, and they are SP.
                  2D video effects and filters, even if they are derived from 3D computations but without physics going underneath, run with SP.

                  And you could still run physics simulations with SP for 3D rendering. The algorithms being used tolerate that. In fact I've been working with a physics simulator for 3D effects and floats (not doubles) were being used. Only in a small fraction of execution time it used doubles in order to not lose precision, but in my opinion that is stupid because the linear solver runs with float values.
                  For some reason, I was under the impression that Professional 3D Renders for the Movie/Tv Industry used DP. I could be wrong. I thought it had something to do with optimizing the ray tracing algorithms.

                  Comment


                  • #19
                    Originally posted by dragorth View Post
                    For some reason, I was under the impression that Professional 3D Renders for the Movie/Tv Industry used DP. I could be wrong. I thought it had something to do with optimizing the ray tracing algorithms.
                    I believe that normally, if you want to optimize, you want to design and tune algorithms to not need DP as much as possible in order to get better performance with the minimum loss in accuracy.

                    Comment


                    • #20
                      Originally posted by dragorth View Post
                      Except that this Titan, the Titan X, isn't. Those 3D professionals need the DP performance of the earlier Titans, and this one doesn't have it. This one is equivalent to the gamer cards in that respect. Those Titans can handle multiple billions of Double Precision, i.e. 64 bit Floating Point numbers, while this card can only handle about 770 million. In fact, this card will be 1/3 the speed of the previous Titan Black in terms of Double Precision. Whereas it's Single Precision numbers go from 5 Billion to 8 Billion operations per sec.

                      So, this card IS a gaming card, compared to even the last year model.
                      DP is mostly for compute (simulations and other special software, typically for research). Most rendering, CAD, video editing etc. don't rely heavily on DP. One of the biggest news in GM200 is the support for half-precision (16-bit float), which is targeted for "neural networks"/"deep learning" which is solely a professional use case. As you know practically no game can utilize the 12 GB of RAM on Titan X, so this card is in no doubt a professional card, not a gaming card.

                      Jen-Hsun Huang also mentioned Nvidia still will provide Titan Z for those who need DP. Unfortunately Nvidia had to sacrifice DP in GM200, but it makes sense since they've pushed 28nm to the limit. So until Pascal arrives, Titan X and Titan Z will remain the (semi-) professional options.

                      Comment

                      Working...
                      X