Cycles Benchmarks – Nvidia update

In-between all other duties here in Blender Institute, Sergey Sharybin is building a 24/7 test farm running Blender’s daily builds to check on various hardware configurations. The tests will be for Cycles renders at first, but will include other tests too.

Purpose is that the daily stats + images will get published online automatically, with a nice website UI and history browsing. (And that’s most of the work, which is a reason to be a bit more patient for now!).

As a sneak peek, here’s the last Cycles stat with Blender 2.79 (master). It includes four new GPUs – provided by Nvidia two weeks ago.

Note that the super fast Quadros are quite expensive! However, prices change all the time – not sure if we should include such information in stats overviews. Just google it.

(Click on image for more info and a live web page)


Get the latest build for testing here:
https://builder.blender.org/download/

-Ton-

Blender 4.3 Beta is here!
Results from Google Summer of Code 2024
VSE Workshop: August 2024
Online Assets Workshop Report

43 comments
  1. Hello, the SI from my work put a config with “quadro p620” for rendering with cycles in 2020
    what could i tell them?
    when i saw this on nvidia site, i’m not convinced to pay for that kind of config with a soo hold cg Thanks for your help.

  2. From experience, the nvidia driver version and settings has to be taken into account too.

  3. Blenchmark runs of fixed (and too small for GPU) tile size, so you could not take it into account.

  4. Could you do a benchmark for the Nvidia Tesla K80 or P100? There is a pretty cheap Google Cloud service offering K80s and P100s to add to your instances. This might be an option to render large scenes etc on if the performance is any good. However, I found a page called blenchmark which shows pretty bad results for these cards…

  5. Theres a option in Blender in the “acene” tab and activate de “Simpify” and

    There is an option on the “Acene” tab and if we activate where it says “Simplify” and keep the default values and only change the value of “AO Bounces Ren” to 3, the render time can reduce almost by half, but everything depends on the complexity of 3D work and the lights.

    This option is not sure exactly what it does but it seems to fiddle with the way the light reflects in the environment and …, simplifies :)

    Try and change the values for your way and watch the render time come down to almost half of what it normally takes.

  6. Mi potete confermate che con imac pro con Radeon pro vega 64 16gb Blender gira bene in opencl render?

  7. New update: tests with the first Titan V (Volta ) and CUDA 9 shows fast hair render.
    Excellent performance numbers:

    10:15.38 – Barbershop
    00:40.20 – BMW (ranked 1)
    02:56.40 – Classroom
    01:28.54 – Fishy Cat (ranked 1)
    02:59.21 – Koro (ranked 1)
    04:18.91 – Pavillion

    For the number crunchers, here are all updated stats from our test lab:
    http://download.blender.org/institute/benchmark171221/latest_snapshot.html
    (Note, click on a GPU/CPU name in top to hide it from the graphs).

  8. Check this from Mike Pan, he found a reason for slower Win10 renders
    http://blog.thepixelary.com/post/167616662857/improving-performance-of-blenders-cycles

  9. News update: Sergey upgraded CUDA version on the Windows buildbot, performance now is better. Please try?

    • koro on 2.79.0-git.d697e3d: 9:58.40
      koro on 2.79 release: 10:12.98

      So not really that much faster.
      Win 10 64bit
      EVGA 1080 Ti Black 11GB

      Going to upgrade my driver and do a clean install, will test again tomorrow and post if results are different.

  10. Trying to have my cake and eat it too after getting over 10 mins on koro on a 1080 Ti Black in Windows 10: I tried installing blender under Ubuntu on Windows and it actually works! I couldn’t run it in a window but console render worked fine once I disabled audio. I wish they had a newer version than 2.69 though as it took 1 hour 31 minutes and 39 seconds to render on a i7-6850K(and didn’t even use them all on the last tile). Doesn’t seem like GPU rendering is an option yet either but maybe a newer version would fix that too(if Windows passes through the GPU to Ubuntu, it’s hard to find any info on what’s supported.

    • Oh, I think it might have taken so long and not used by CPU cores efficiently because I tried to render the koro_gpu.blend file to use my GPU(which 2.69 doesn’t have the CUDA core for). Unfortunately, they don’t currently support GPUs in Ubuntu on Windows but it is their most requested feature.

      koro_cpu.blend appeared to use my cores properly. Here’s the stats in case anyone is interested in Windows vs. Ubuntu on Windows Blender 2.69 in console/command line:

      Ubuntu on Windows Blender 2.69 64bit, koro_cpu.blend: 55m:07s
      Windows Blender 2.69 64bit, koro_cpu.blend: 1h:13m:53s

      So even the Ubuntu on Windows version is 18 minutes faster than the native Windows binary, which is a real interesting result. This is 2.69 though, so this is more of an experiment than a real useful result for modern Blender. Need to see if it’s possible to compile 2.79 next.

    • I managed to compile 2.79 on Ubuntu for Windows and it’s about a minute faster than the native Windows binary on CPU. Real interesting.

  11. What happened to the Vega 56 benchmark?

    • We replaced it with vega 64. The 56 is now in use by Eevee developer Clement.

  12. Regarding render time differences between this benchmark and Windows, there’s a lot of factors here. Rendering in the background vs. in the UI, using the GPU for display or not, lack of compute preemption support in the Windows driver, CUDA toolkit 8 or 9, faster BVH building on Linux, small tile sizes now often rendering a bit faster than large ones, etc.

    Anyway, we’ll investigate this, the recent CUDA rendering optimizations have mostly been tested on Linux so far.

  13. Like I said I have 2x Nvidia 1080’s, and I’ve tested both and get consistent results, there Gigabyte and where purchased new from Newegg. Tested with and without overclocking, BMW is 2:33 without OC, and 2:16 with OC. So definitely beneficial. I’ve also run the official Classroom Benchmark, and results are same, Classroom Render time: 5:48 sec, about 32% slower than the Institutes render times.

    Yes I’m on Windows 10 Pro, currently don’t have a boot drive for Linux but will test that when I set one up. Hard to believe that Linux drivers make that much difference though? I mean 32% is the same percentage difference between the Nvidia 1080 and the 1080Ti in performance rendering.

    Also really like the idea of a website or page for official bench-marking results. Definitely will help where I can with this, as I feel it would be a great asset to Blender users.

    • I checked, the machine we used for testing had a very fast and beefy xeon cpu. That matters too. We’ll update all tests on systems with similar i7s.

      • Okay I spent yesterday installing Linux and drivers etc, and running benchmarks on both my GTX 1080’s no OC. And to my surprise I rendered the BMW scene in 1:40! compared to 2:33 on Windows 10. So that’s the same as the institutes time, I also ran the Classroom scene and times matched the institutes as well. This means that rendering on an Nvidia GPU is up to 50% faster on Linux vs Windows 10! There are definitely some issues up, possibly on Nvidia’s Windows drivers.

        But due to the huge difference in performance that Nvidia cards have between Linux and Windows, and the fact that the very large majority of users run Windows 10. I think the Institutes benchmarks should be run on Windows 10, also since AMD cards show no difference between the 2 operating systems, it currently makes OpenCL cards a much better value, If the above charts showed the 30-50% slower rendering times for Nvidia on Windows, AMD cards would be the better option by far.

        Also thought I’d mention that Threadripper was slightly faster on Linux as well, coming in at 5:54 in the Classroom scene.

        • Wow. That’s interesting. We test OpenCL in both Linux and WIndows, for CUDA we stick to Linux here. Always had good performance with it.
          It’s something we’ll keep track of, would also know from others if they have similar experiences. I don’t agree we “should” only test Windows here though. Rather the opposite. It’s an open source studio here.

          Note that historically, Linux always performed better for rendering. It’s a good server OS for a reason.

          • Agreed on that, but I think bench-marking on both would be very beneficial, most people using the charts to decide what GPU to get for Blender will likely be using Windows 10, and probably be confused as to why there results are so different due to rendering on a different OS.

            Also strange that OpenCl is identical on both Linux and Windows, and Cuda can be up to 50% slower, will see if I can find others to test this as well.

          • What version of linux do you recommend for Blender?
            In your videos I’ve seen Xubuntu linux with XFCE. Can you confirm it?

            If performance are better than windows I’ll install a dual boot system only for Blender.

  14. We test by rendering twice (only registering 2nd render time), and we render in command line (no UI). And in Linux… I’ll ask Sergey tomorrow if he has ideas that explains the difference.

    Note: the Threadripper has a big performance issue with some part of the render initialization – we cannot find out what (don’t have such system here yet, AMD works on it).

    • Oh wait! Our numbers are not with 2.79 release, it’s with the current 2.79 “master” …
      Get the last build here and check? https://builder.blender.org/download/

      (Will fix the header line of the stats page)

      • Okay good to know! will run the benchmarks again and see if I can get more comparable results.

      • So results with the latest 2.79 Master build look more in line with what they should be on Threadripper, with the Classroom taking 6:06 and BMW 1:54. Also my Radeon Pro Duo matches the Institutes in render times after excluding the render kernal load time.

        But the GTX 1080 is off quite a bit, and I’ve bench-marked 2 overclocked 1080s now getting the same results, latest drivers, and not using them as a display, render times are usually 30%+ more for me, with the BMW Gpu benchmark taking 2:16 compared to the 1:43 on your charts.

  15. Yeah my numbers don’t really add up either, using Blender 2.79 and the Blender Institute prepared scene files, the CPU Classroom Benchmark takes 8 min on an OC 3.9Ghz 1950x Ryzen Threadripper processor, compared to just 9:50min on a Ryzen 7 8-core 1800x? I’ve done some of my own testing comparing Ryzen 7 to threadripper and its been about what you would expect with the 16 core Threadripper chip rendering twice as fast as Ryzen 7.

    • Update with my GTX 1080 results on the GPU classroom benchmark, testing 2 different 1080’s OC I’m getting 6:00+ render times on Blender 2.79, compared to 4:22 on the Blender Institutes GTX 1080? Something must not be right that these numbers aren’t matching up.

  16. I’m a bit surprised…I get only 2.16 minutes on my Asus ROG Strix 1080 TI OC edition compared to the stats claimed 1.12 minutes. How is that even possible?

  17. Hmmm, I have benchmarked a good number of cards a few weeks ago with the same version of Blender (v2.79 release) and the results are a bit all over the place compared to this. What was the tile size for dGPUs – 256×256? Unfortunately, only the RX 480 is the same data point. Any ideas what may be behind the differences?
    bmw – 3:48 vs 3:00
    classroom – 8:32 vs 5:47
    pavillon_barcelona – 11:30 vs 12:29
    http://boostclock.com/show/000149/blender_2-79-AMD-OpenCL-dGPUs.html

  18. The GTX 1060 and 1080 seem to struggle quite a lot on the Koro scene. Something to do with the hair I guess.

  19. Great! It’s often quite tedious to find proper, relevant, trusted benchmarks.

  20. I know the purpose is to compare the different cards, though it might be interesting to know how many samples you used on each scene to have a global idea on the render time/quality?

  21. Where’s the RX Vega 64? With those now at $499 US compared to the $750 for the 1080Ti never mind the thousands for the Quadro it sure would be nice.

    How about testing against the SSG?

    If you’re going to show the Quadro 6000 (GP100) you might as well show the Radeon Pro Vega WX 9100.

    Or do you not think AMD will offer to gift you that and Threadripper? After all, they did the work on OpenCL for you, for all of us.

    You’d think you’d want to highlight their products in the professional line.

    • We will publish a RX Vega 64 test soon. It needs a new system here with beefy power supply.
      AMD is working on sending us a Threadripper. Demand is high apparently.

  22. regarding of the price it could also get the prices for the cards averaged from google search on that day, so you can have a real price/perfomance comparison
    i know it might be too much to ask, but :D

In order to prevent spam, comments are closed 7 days after the post is published.
Feel free to continue the conversation on the forums.