Announcement

Collapse
No announcement yet.

NVIDIA Announces The GeForce RTX 40 Series With Much Better Ray-Tracing Performance

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

  • #61
    Originally posted by WannaBeOCer View Post

    Nvidia no longer makes GPUs just for gamers. They target multiple different markets with their consumer GPUs. R&D cost have been going up about 35% every year as well. RDNA will kick ass in gaming aside from that it will be a disappointment just like RDNA1/2. RDNA2 was the worst at launch since they priced their pure gaming GPUs about the same price as Nvidia’s Ampere.
    To take this point a little further: I don´t think NVIDIA actually makes consumer GPUs. They make variants of one or two GPUs (e.g. Lovelace & Hopper) and wrap products around those, some of which are targeted towards the consumer. As an example, the H100 is an NN/ML accelerator targeted towards huge AI workloads, whereas AD10x seems to be for every other market (GeForce RTX for games, RTX <whatever> for the professional workstation market and the A40 & L40 for the data center workloads such as VDI/virtual workstation, video encoding, ¨small¨ AI/ML workloads).

    Having thought about it, this might explain why NVIDIA´s gaming GPUs are getting so big & power-hungry - they are actually professional/data center GPUs repurposed for PCs and given technologies to make use of all that silicon (raytracing, DLSS) - notice that with Ampere, the higher-end SKUs didn´t really shine until 4K. Compare this to AMD´s RDNA and CDNA architectures, each of which is aimed specifically at a given market.

    Comment


    • #62
      Originally posted by parityboy View Post

      To take this point a little further: I don´t think NVIDIA actually makes consumer GPUs. They make variants of one or two GPUs (e.g. Lovelace & Hopper) and wrap products around those, some of which are targeted towards the consumer. As an example, the H100 is an NN/ML accelerator targeted towards huge AI workloads, whereas AD10x seems to be for every other market (GeForce RTX for games, RTX <whatever> for the professional workstation market and the A40 & L40 for the data center workloads such as VDI/virtual workstation, video encoding, ¨small¨ AI/ML workloads).

      Having thought about it, this might explain why NVIDIA´s gaming GPUs are getting so big & power-hungry - they are actually professional/data center GPUs repurposed for PCs and given technologies to make use of all that silicon (raytracing, DLSS) - notice that with Ampere, the higher-end SKUs didn´t really shine until 4K. Compare this to AMD´s RDNA and CDNA architectures, each of which is aimed specifically at a given market.
      I disagree, Turing, GA102 and Lovelace are general purpose GPUs aimed at consumers. Nvidia split their architectures a while back with the release of GP100 vs GP102. Then introduced their tensor accelerator cores(tensor cores) in Turing and continued their split architectures with TU102 vs V100, GA100 vs GA102 and now H100 vs AD102. Gamers need to stop thinking they're the only individuals that utilize GPUs. Just because an individual utilizes them for parallel computing or content creation doesn't mean they should have to shell out $3-5k for a GPU.

      AMD has been two generations behind since 2010, they eventually followed Nvidia's revolutionary computing architecture Fermi with the release of Tahiti. Then finally releasing a computing platform ROCm with Vega. From leaks it appears they will add tensor accelerators similar to matrix cores of CDNA in RDNA3 which is great news. Nvidia uses them specifically for ray tracing denoising and upscaling but tensor accelerators can be used for physics simulation, character locomotion and audio to facial animations in regards to games.
      https://www.youtube.com/watch?v=8oIQy6fxfCA

      Intel also showed off an AI utilizing Nvidia's Tensor cores in a RTX 3090 to make GTA V photorealistic.
      https://www.youtube.com/watch?v=P1IcaBn3ej0​
      Last edited by WannaBeOCer; 21 September 2022, 11:20 PM.

      Comment


      • #63
        If DLSS is boosting up to 4x, that means they are rendering 1080p and upscaling it to 4K, not sure I want that..

        Comment


        • #64
          Originally posted by WannaBeOCer View Post
          I disagree, Turing, GA102 and Lovelace are general purpose GPUs aimed at consumers. Nvidia split their architectures a while back with the release of GP100 vs GP102. Then introduced their tensor accelerator cores(tensor cores) in Turing and continued their split architectures with TU102 vs V100, GA100 vs GA102 and now H100 vs AD102.
          Ampere & Lovelace GPUs are used in the A40 and L40 accelerators respectively, which are clearly not consumer-level products - they are targeted to the datacenter. I believe the GA100 and GA102 GPUs are two variants of the same basic architecture, Ampere - in the case of the GA100 I believe that the parts of the micro-architecture specifically built for visualization were removed (or not added in the first place) and more matrix accelerators/tensor cores where added instead (I´m simplifying here). On the other hand Hopper was built specifically for AI in the datacenter and therefore never had any visual elements in its micro-architecture.

          Originally posted by WannaBeOCer View Post
          Gamers need to stop thinking they're the only individuals that utilize GPUs. Just because an individual utilizes them for parallel computing or content creation doesn't mean they should have to shell out $3-5k for a GPU.
          Can you clarify what you are saying here? I ask because the two statements appear to be contradictory.

          Comment


          • #65
            What about supported codecs?

            Comment


            • #66
              Originally posted by MorrisS. View Post
              What about supported codecs?
              AV1 decoding was already supported by the RTX 30 series.

              The RTX 40 series supports 8K 60fps dual stream AV1 hardware encoding.

              Comment


              • #67
                Originally posted by carewolf View Post
                If DLSS is boosting up to 4x, that means they are rendering 1080p and upscaling it to 4K, not sure I want that..
                Maybe read about how DLSS 3.0 works first?

                Comment


                • #68
                  Originally posted by birdie View Post

                  Maybe read about how DLSS 3.0 works first?
                  How do you think an upscaling algorithms achieves speedups? Do you think doing extra work, somehow makes the base rendering faster? It has to base work + the time it takes the DLSS algorithm, so with 4x speedup, it HAS to do at maximum a quarter of work before DLSS is applied, there is no other way.

                  Comment


                  • #69
                    Originally posted by carewolf View Post
                    How do you think an upscaling algorithms achieves speedups? Do you think doing extra work, somehow makes the base rendering faster? It has to base work + the time it takes the DLSS algorithm, so with 4x speedup, it HAS to do at maximum a quarter of work before DLSS is applied, there is no other way.
                    I am sorry to say birdie is right that you do need to read how DLSS 3.0 works. DLSS 3.0 added keyframe rendering.

                    So yes you are rendering at 1080 for the mid frames but then you are rendering 4K keyframe every so often and comparing to what the AI upscale generated from the 1080.

                    I can possibility of some really wacky artifacts with DLSS 3.0. I am not sure if the 4K keyframes will be ever directly shown on screen or will be just used for self tuning of the AI upscale with DLSS 3.0. When I say wacky this could mean feed the same data by a playback program into GPU and get two very different outputs..

                    There is one thing here absolutely DLSS 3.0 is not designed to upscale existing non modified games.

                    So DLSS doing 1080p to 4K will be mix rendering in 1080p and 4K. Way less 4K frames. Of course this brings interesting problem. GPU memory usage back into play. Rendering 4K means you need 4K textures loaded and rendering 1080p means you need 1080p textures loaded. I feel sorry for the Nvme drives.

                    Comment


                    • #70
                      Originally posted by parityboy View Post

                      Ampere & Lovelace GPUs are used in the A40 and L40 accelerators respectively, which are clearly not consumer-level products - they are targeted to the datacenter. I believe the GA100 and GA102 GPUs are two variants of the same basic architecture, Ampere - in the case of the GA100 I believe that the parts of the micro-architecture specifically built for visualization were removed (or not added in the first place) and more matrix accelerators/tensor cores where added instead (I´m simplifying here). On the other hand Hopper was built specifically for AI in the datacenter and therefore never had any visual elements in its micro-architecture.



                      Can you clarify what you are saying here? I ask because the two statements appear to be contradictory.
                      I'm aware of the passively cooled server variants, AMD also has the Radeon Pro V620 based on RDNA2 along with the Radeon Pro W6800 for workstations. GA100 and GA102 SMs look completely different. GA102 has 2x FP32 Processing, GA102 has 6 MB of L2 cache compared to GA100's 48 MB along with FP64 units. As mentioned AMD is just 2 generations behind and we'll see them catch up slowly in regards to ML on their consumer GPUs.

                      My argument is that Geforce cards are still consumer cards but aimed at additional markets, content creators with Studio drivers along with the continued support of CUDA. They added better encoders and tensor cores for these consumers. A student/hobbyist that's into content creation/ML is still a consumer not a professional and doesn't require ECC, tested drivers for professional workloads, SR-IOV and etc.
                      Last edited by WannaBeOCer; 22 September 2022, 12:53 PM.

                      Comment

                      Working...
                      X