More Darktable GPU/CPU Benchmarks - 27 Different Setups

[WorksAsIntended]

Very interesting, thank you!
About the Intel HD problem: for some reason it is also necessary to install intel-opencl in addition to beignet to make the intel integrated graphics work (7th gen).
Here some results on my t470s with 7500U (ntel HD620):

https://openbenchmarking.org/result/1809022-AR-SDF47371953

The integrated graphic is about 1/3 faster than cpu only. Still to slow to use darktable regulary.

[taxi_bs]

It would be helpful to post the full output of darktable-cli. Then we could see the needed time of CPU and GPU for _every_ module. The big difference between the two i5 are only possible at either denoise or equalizer module (or both together). No idea about testcase 2 and 3, because i don't know which modules are used there.

[sdack]

Any idea why a i5-6500 is nearly twice as fast as a i5-6600k?

Also GTX 980Ti being 10x faster than a GTX 980? Obviously can these numbers not be used in any meaningful way outside this particular test and setup, making the benchmark about as remarkable as a doorstop on a motorway.

A little less numbers and more factual reporting would have made it a much better article.

[boxie]

Thank you Michael for all the testing, i think darktable and opencl deserve this attention.
if you want to burn you cards in a heavier test: You could take a BIG raw like one from Hasselblad medium format:

http://www.rawsamples.ch/raws/hasselblad/h3d-39ii/RAW_HASSELBLAD_H3D39II.3FR

and then apply my xmp-file with this raw. Maybe the ranking is the same, maybe not. Looks like it takes 6 times longer than the raw from my Samsung on a GTX 1060. Which is a lot.

Finally let's hope that we get soon some numbers with the open source drivers.....

Michael this!

[finalzone]

Benchmark on ASUS X550ZE using Radeon R7 M265DX seen as Radeon R6 and running on Wayland session:

https://openbenchmarking.org/result/1611282-LO-1611284TA88&obr_sor=y&obr_hgv=Radeon+R7+M265DX

[taxi_bs]

Thank you Michael for all the testing, i think darktable and opencl deserve this attention.

Some words to the benchmark:
It's not a boat, it's a ship on Lake Lucerne. And anyone who uses this photo for testing should spend his/her next holiday in Switzerland ;-)

I activated usual modules in the xmp-file, just added denoise and equalizer as some bonus of heavy gpu-workload. So i think it is a really good example of what opencl can do here for a normal photographer. Of course many people never use these modules, and a full export with only some little white-balance is much faster. In that case, you won't see much difference with or without gpu.

Just keep in mind that this benchmark is about export. During interactive work, the gpu also helps very much in faster applying modules or zooming in/out the picture. So _any_ of the tested gpu is very good, some are just better. Personally i am a bit surprised about the big gap between GTX 1050Ti and GTX 1060.

For those who want to investigate the modules in darktable itself, here are the two files:

http://www.mirada.ch/bench.SRW

http://www.mirada.ch/bench.SRW.xmp

darktable-cli bench.SRW test.jpg --core -d perf -d opencl
or
darktable-cli bench.SRW test.jpg --core --disable-opencl -d perf

Its a Samsung aps-c-raw, by the way.

Michael, if you want to burn you cards in a heavier test: You could take a BIG raw like one from Hasselblad medium format:

http://www.rawsamples.ch/raws/hasselblad/h3d-39ii/RAW_HASSELBLAD_H3D39II.3FR

and then apply my xmp-file with this raw. Maybe the ranking is the same, maybe not. Looks like it takes 6 times longer than the raw from my Samsung on a GTX 1060. Which is a lot.

Finally let's hope that we get soon some numbers with the open source drivers.....

[hiryu]

Maxwell, I don't have the Pascal.

Then that's extremely impressive that it outperformed everything else, including the Pascal cards. Maybe it's the extra memory?

[schmidtbag]

Any idea why a i5-6500 is nearly twice as fast as a i5-6600k?

Seems to me Darktable is poorly optimized with OpenCL. It also doesn't really make sense how the RX 480 in one test is over twice as fast as the Fury. Though some of the Nvidia results are weird, I think a lot of those can be blamed on driver issues, but most of these results seem weirdly out-of-place.

[darkbasic]

By the way, *VERY* interesting benchmark Michael. But it would be interesting to see Beignet vs CPU on Iris.

[darkbasic]

Latest Beignet works for me on Broadwell:

*GPU*: [dev_process_export] pixel pipeline processing took [B]*64,495*[/B] secs (14,497 CPU)

*CPU*: [dev_process_export] pixel pipeline processing took [B]*43,410*[/B] secs (168,420 CPU)

model name : Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz
model name : Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz
model name : Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz
model name : Intel(R) Core(TM) i7-5600U CPU @ 2.60GHz

00:02.0 VGA compatible controller: Intel Corporation HD Graphics 5500 (rev 09)

Linux arch-laptop 4.8.10-1-ARCH #1 SMP PREEMPT Mon Nov 21 11:55:43 CET 2016 x86_64 GNU/Linux