The new Cycles GPU 2.72 Benchmark

Finally got my new Quadro K2200 to work with Blender.

Here is my result:

Gentoo Linux 64 bit (kernel 3.12.13)
Intel i5-650 @ 3.2 GHz
NVIDIA Quadro K2200

26:17.05 (Tiles: 360x810)
23:34.25 (Tiles: 240x270)
22:56.13 (Tiles: 256x256)

Silly question maybe but did you install CUDA 6.5?

why no hair why no SSS bad benchmark.

Yep, I even tried installing it separately, as some of the drivers in the Mint repositories don’t have CUDA included. I’m going to give Mint one last try tonight, and see how a purge of the drivers and a reinstall from the Edgers PPA will work.

Ok I got the card working. :slight_smile: I had to remove the 980 from the computer, setup the driver with just the 460, then replace the 980 and it’s working. I ran the benchmark and this was the result:

Render time: 07:53.87
OS: Linux Mint 17

My card is a MSI GTX980, running at 1328 Mhz and memory transfer rate of 6008 Mhz. Even with the performance as it is, it’s still 5 to 6 times faster than my 460. :smiley:

Thanks for your help Anthony.

Jason

Glad you got it running.

Wierd that 980 is 2-3min slower than 780Ti.

The 780Ti has more cores and a faster memory interface than the 980.

It might improve with driver updates and/or Cycles enhancements?

OS: Ubuntu 14.04
GPU: GTX 580
Render Time: 13:09.24

Tested GTX 770 and 670 combined in addition to my previous results.

Windows 7 x64
GTX 770 + GTX 670
05:48.57

I must agree that the tile size is not optimal indeed, even for a single GPU.

New entries, I hope all in this time.
@Grimm, are you using the build from the post > http://www.blenderartists.org/forum/showthread.php?350975-The-new-Cycles-GPU-2-72-Benchmark&p=2743138&viewfull=1#post2743138

The native kernel_sm_52.cubin is compiled for GTX 970/80 and should be faster than renamed sm_50 kernels.


Cheers, mib
EDIT:Hehe, forgot the whole picture. :slight_smile:

Hi at all,

here my result…

Windows 8.1
Asus GTX 770
Render time: 10:37:68


I think that with this much input it gets hard to get the big picture from the sheet. As you have all the data in excel couldnt you easily group graphic cards with same model, take average and then generate a tower graph which shows graphically how fast cards are compared to each other? :slight_smile:

Would be cool anyway.

@jjjuho, I add DL link on first page to my Libre Office file, add xls file too.

Cheers, mib

Yep and I copied the new 52 kernels in to the cycles lib directory. I re-ran it to see if it would come out different, and it did. 7:50.98 What this tells me is that the usual methods for setting up the kernel don’t really work well for maxwell if the times can vary so much. At the very least some tweaking would probably help. I’m seeing this same variability in Octane too.

I ran the benchmark again and included the 460 for grins and giggles. 6:22.18 The 460 improved the render by over a minute and a half. Of course the 980 had finished 3 of the tiles before the 460 was done with one.

@mib
will that kernel_sm_52.cubin work for my gtx 750 ti, i am running of the default kernel, never replaced anything…

Windows 8.1
Gigabyte GTX 780 GHz (GV-N780GHz-3GD)
Render time: 06:17.09

Windows 8.1 64 bit
MSI 970 @ 1527Mhz core and 4037 Mhz memory
OscarM blender build
Time: 07:58:85

For comparison
Official blender build
Time: 08:15:84

Hi Davin101, the GTX 750Ti is a sm_50 card and is official supported in Blender 2.72.
@Chuk_Chuk, it is only 3.5% but interesting, thanks.

Cheers, mib