7 Replies Latest reply on Mar 28, 2017 12:39 AM by Taloot Abdullah

    Visualize - GPU CUDA core count

    Ben Zuckerman

      Hi There, We're looking at upgrading our GPU's from two GTX 780 Ti to two GTX 1080's or 1070's.  From what I've read and pass experience from Bunkspeed days CUDA cores are one area you want to maximize for your $$. As expected when doing a comparison the 1080/1070 will give us a significant jump in memory and speed but was surprised they have less CUDA's than our 780's, Confusing! Anyway assuming we'll still get a significant increase in rending speed if we upgrade to the 1080's can anyone give us an idea (approx) by how much. Also would you have an idea in rendering speed difference between the 1080 & 1070.

      Note sure if this would affect the response but we render anything from small sized consumer products to larger industrial equipment.

      Thanks for your time,

      Ben

       

      BTW wouldn't it be awesome from a users perspective if there was some sort of visualize calculator (like an online percentage calculator)  that you could input your GPU specs (or even better model no.) and maybe a typical render you do and it can give you an idea of speed/time differences?

       

        • Re: Visualize - GPU CUDA core count
          Paul Salvador

          The program has the ability to recognize the gpu and cpu... so there has to be a way to collect/share the information for all to compare?

          They know what performs better...  maybe it's a privacy issue... so.. if there was a option for the user to click "Agree" to send their system specs during/after a test scene... it would be helpful, imho.

          • Re: Visualize - GPU CUDA core count
            Ryan Navarro

            To get an idea of the relative performance of GPUs I like to use Passmark benchmarks as a guide. It lets you search for and compare particular cards if desired. The cards you mentioned are all near the top of the list so should be easy to compare.

            PassMark Software - Video Card Benchmarks - GPU Compute Video Cards

             

            The key thing is knowing what benchmark to look at. In this case you'd want to look at GPU Compute. These results could be very different for gaming benchmarks for instance. But Visualize is using CUDA cores for GPU compute so that should be the most relevant comparison.

             

            The Passmark benchmarks aren't using CUDA language specifically, but I found some CUDA benchmarks on another site and they appear to show similar trends

            CUDA Benchmarks - Geekbench Browser

             

            Comparing CUDA cores is tough to do across generations of cards, since there may be architectural changes affecting performance per core. In this case the clock rates on the newer GPUs are a lot higher but there are less cores. So similar to CPUs where you can't go by "just" the Ghz, it's better to look at benchmarks. The higher VRAM on the newer cards should also handle larger problems. That's one of the benefits of the Quadro cards - they have a lot more VRAM, along with the officially certified drivers.


            If anyone has direct experience with these cards then I would definitely trust that over the benchmarks.

             

            Maybe we can get a benchmark thread going for Visualize like we have for SOLIDWORKS?

              • Re: Visualize - GPU CUDA core count
                Ben Zuckerman

                Ryan thanks so much for your feedback - your links are helpful and make it easier to quantify the upgrade.

                Happy to hear from anyone with direct experience using the gtx 1080 or any of the 10 series.

                A Visualize GPU benchmark thread is a great idea!  Bunkspeed use to have one along with a standard render to benchmark against.  I found it really useful for GPU upgrading and you could cross reference your setup to confirm it was operating within range or under preforming.

                  • Re: Visualize - GPU CUDA core count
                    Ben Zuckerman

                    UPDATE: So we took the plunge and upgraded one of our machines with dual GTX 1080 (unfortunately did it a few days before seeing the 1080 Ti release)  Did a couple of rendering speed comparison with a machine running dual GTX 780 Ti and only found a 20 - 25% speed increase.  Both machines have identical CPU, RAM, HD, power supply, etc with exception of their GPU's. Rendering setup was identical (passes, size, GPU only, etc). No other programs running etc.  SLI is enabled on both.  Does 20 - 25% increase sound right or are there any tips for increasing the speed on the 1080's?

                • Re: Visualize - GPU CUDA core count
                  Taloot Abdullah

                  ben you didn't specify your machine cpu for me i have GTX 1070 + 1060 3g about 40 times faster than my cpu which is i5 6600k

                   

                  i think running 2 gtx1080 will bottleneck by your cpu, also on my machine overclocked to 4.4ghz  the cpu load while rendering in both gpus about 60-65%

                   

                  your cards can make up to 18 tflops , and mine about 9.5 so i assume you bottelneck by cpu