HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Announcement of new releases, bugs, support, suggestions

Moderator: Stas Yatsenko

pwnell
Posts: 32
Joined: 16.12.2012 21:40

HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 29.12.2017 06:25

I have just received my new iMac Pro 10 Core system. I also have a Mac Pro 2013 with 6 Cores. How is it that rendering a stack of 268 DNG files using method B, and Adobe DNG converter, takes 8min 55sec on the 6 core Mac Pro but 11min 12sec on the 10 Core iMac Pro with its much newer CPU (even though base clock is 500MHz slower)?

Cinebench gets 2045 points on the iMac Pro, and only 950 on the Mac Pro for multi core performance.

That is not even speaking about my Windows machine, 6 Core 6850K CPU overclocked to 4.4GHz that runs the same test in under 3min 39sec. What is going on here? Is Helicon Focus not optimized for 10 cores? It showed 1900% CPU usage most of the time...

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 29.12.2017 16:48

We have recently made significant improvements in both per-core processing speed and multi-core optimization. We are now getting good results from a 24-core Intel and 16-core AMD CPUs that we used for testing so the next version of Helicon Focus should work much better. The only problem is that the upcoming updated version of the program is still work in progress. It's already working on Windows with only some UI changes not finished yet, but the Mac version is not functional at the moment so I can't even give you a preliminary build to try right now, unfortunately.

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 29.12.2017 20:48

Might I get a copy of the Windows version? I am running Parallels and that would be so awesome!

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 02.01.2018 20:54

Sure: https://drive.google.com/open?id=1HN-ak ... St6YC6_jFZ
Please note that this version is far from final and has some problems, but should generally be working.
Also note that you will lose some performance to the virtual machine layer, but you should still get a decent boost out of your multi-core CPU.

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 02.01.2018 21:05

Thanks, trying it now. What is the difference between the two Method B's? (depth map and legacy)

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 03.01.2018 11:53

Depth map is a new experimental method. You can compare it with the old Legacy method.

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 22.02.2018 19:07

I thought I'd update you with the performance differences I saw.

I downloaded 6.8.0 and stacked 268 DNG images each 20.4MB and 20MP using method B, as well as the unreleased new version (On a mac stacking DNG is horribly slow but I do it for obvious reasons - stacking TIFF is way faster. On Windows stacking DNG is about twice as fast as the mac for similar hardware - I think their Windows version is much better optimized - that said, below tests were Windows in Parallels on iMac Pro only).

On my 10 core iMac Pro with 128GB RAM the timings were as follow (on Windows 10 in Parallels with 10 cores assigned to VM, not 20 hyper threaded cores as that killed performance):

Old HF Windows 6.8.0: 8min 22sec (First Run), 4min 57sec (Second run)

New HF Windows 7.0.0 (B legacy mode): 4min 45sec (First Run), 3min 42sec (Second run)

Old HF on iMac Pro in Mac (6.7.2 Mac version):16min 15sec (First Run), 11min 13sec (Second run - it caches the DNG conversions I think)

Why is it so slow on the Mac? The old Windows version is twice as fast as the old mac version even when virtualized and not having access to the 10 additional hyper threaded cores and only 16GB RAM.

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 23.02.2018 10:31

I'm fairly sure it's slower on Mac due to differences in Adobe DNG implementation. You're benchmarking the Adobe DNG decoder + Helicon Focus. If you were to benchmark Helicon Focus alone, you'd have to process TIFF or JPEG files, and I'm pretty sure you wouldn't see such a difference then.

Also, in our testing, hyper-threading always yields some performance boost, at least on Windows. The fact that you've had better results with 10 cores/threads vs. 20 is probably a quirk of the virtual machine.

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 02.03.2018 01:25

Hi Stas,

I did some additional testing. Please see below. Very weird is that Parallels using 18 threads (leaving 2 for host machine) is twice as slow as Parallels with 10 threads (matching the actual core count) using your new version 7.0.0. I tried with and without adaptive hypervisor turned on and it made no difference.

Also, it is clear from below tests that HF on Windows is twice as fast as HF on Mac when using TIFF files. I would love it if Mac could perform better - this is on identical hardware, with Windows penalized by virtue of being virtualized, and still outperforming the host iMac Pro.

Canon 7D Mark II files, 20MP
10 Core iMac Pro, 20 HT Cores, 128GB RAM

Mac HF 6.7.2 Method B, Adobe DNG Converter 10.2, 20 HT Cores
50 x TIFF: Run 1: 1min 55sec (one pass), Run 2: 1min 55sec (one pass)
50 x DNG: Run 1: 2min 58sec (one pass), Run 2: 2min 6sec (one pass)

Win HF 6.8.0 Method B, Adobe DNG Converter 10.2, 10 Virtual Cores
50 x TIFF: Run 1: 50sec (one pass) , Run 2,3,…: 1min 38sec (two passes)
50 x DNG: Run 1: 1min 6sec (one pass), Run 2,3,…: 59sec (two passes)

Win HF 7.0.0 Method B (Legacy), Adobe DNG Converter 10.2, 10 Virtual Cores
50 x TIFF: Run 1: 35sec (two passes), Run 2: 30sec (two passes)
50 x DNG: Run 1: 50sec (two passes), Run 2: 35 sec (two passes)

Win HF 6.8.0 Method B, Adobe DNG Converter 10.2, 18 Virtual Cores
50 x TIFF: Run 1: 1min 2sec (one pass) , Run 2: 59sec (one pass)
50 x DNG: Run 1: 1min 15sec (one pass), Run 2: 1min 9sec (two passes)

Win HF 7.0.0 Method B (Legacy), Adobe DNG Converter 10.2, 18 Virtual Cores
50 x TIFF: Run 1: 1min 5sec (two passes), Run 2: 51 sec (two passes)
50 x DNG: Run 1: 1min 17sec (two passes), Run 2: 50 sec (two passes)

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 02.03.2018 11:00

You are comparing 6.7.2 (which is slower) on Mac with 7.0.0 (which is faster) on Windows. Perhaps you've made other measurements that you didn't include here and that do support your conclusion, but what I see here is Focus 7 being faster than 6, not Windows being faster than Mac. You need to compare apples to apples.

Our fresh build of Focus 7.0.0 for Windows: https://drive.google.com/open?id=1ns9Ge ... bb_y9kT9mL
And for Mac: https://drive.google.com/open?id=11EFzl ... 2qaeTapYW5

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 02.03.2018 18:55

Hi Stas,

I think you misread my test results. Look at the first two sets of tests:

Mac HF 6.7.2 Method B, Adobe DNG Converter 10.2, 20 HT Cores
50 x TIFF: Run 1: 1min 55sec (one pass), Run 2: 1min 55sec (one pass)
50 x DNG: Run 1: 2min 58sec (one pass), Run 2: 2min 6sec (one pass)

Win HF 6.8.0 Method B, Adobe DNG Converter 10.2, 10 Virtual Cores
50 x TIFF: Run 1: 50sec (one pass) , Run 2,3,…: 1min 38sec (two passes)
50 x DNG: Run 1: 1min 6sec (one pass), Run 2,3,…: 59sec (two passes)

First test set was performed on a Mac, using the latest production release of HF 6.7.2 for Mac. Second test set was performed on Windows in Parallels on the same Mac hardware and 10 virtual cores allocated on HF 6.8.0 - the latest production release of HF on Windows, which is NOT the new 7.0.0.

1. For this test, the "old" HF 6.8.0 on Windows was 2.3 times FASTER than the Mac when using TIFF on the first run.
2. For this test, the "old" HF 6.8.0 on Windows was 1.2 times FASTER than the Mac when using TIFF on the second and subsequent runs. I am not sure why HF on Windows for TIFF performs one pass on the first test execution, and two on subsequent tests as the image looks identical. Since I usually run a stack only once, for me for TIFF I see a 2.3x increase in performance for old HF on Windows over Mac.
3. For this test, the "old" HF 6.8.0 on Windows was 2.7 times FASTER than the Mac when using DNG on the first run.
4. For this test, the "old" HF 6.8.0 on Windows was 2.1 times FASTER than the Mac when using DNG on subsequent runs. I get that you feel Adobe's DNG is slower on Mac - but after I assume to be a cached TIFF on a second run, old HF on Windows still is more than twice as fast than the Mac.

All these have not even looked at 7.0.0 which I detailed after these two test results. So I believe I am indeed comparing apples to apples. Please provide me your thoughts on this.

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 05.03.2018 11:55

Not that I misread your results, more like didn't pay enough attention to that section. 6.7.2 is not that much different from 6.8.0, but still, these are different versions and should not be directly compared. Also, these are old versions. Focus 7 has different performance characteristics, and there's simply little sense in benchmarking the old versions that are about to become obsolete. That's why I encouraged you to test Focus 7 instead.

P. S. The one pass / two pass difference is because of memory consumption. Single pass requires all the images to be loaded into RAM, and if not enough RAM is available, two passes are necessary. And due to improvements in image caching, you may not see a second pass in Focus 7 (or you may. Hard to tell, it depends on several factors).

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 05.03.2018 19:55

Ah that makes a lot of sense, thanks.

I only allocated 16GB of RAM to the VM so that may have been it.

Will test the new 7.0.0 versions.

pwnell
Posts: 32
Joined: 16.12.2012 21:40

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by pwnell » 05.03.2018 21:54

Not sure if I am missing something but HF 7 for Mac seems stupendously fast:

Old:
Mac HF 6.7.2 Method B, Adobe DNG Converter 10.2, 20 HT Cores, 128GB RAM
50 x TIFF: Run 1: 1min 55sec (one pass), Run 2: 1min 55sec (one pass)
50 x DNG: Run 1: 2min 58sec (one pass), Run 2: 2min 6sec (one pass)

New:
Mac HF 7.0.0 Method B (legacy), Adobe DNG Converter 10.2, 20 HT Cores, 128GB RAM
50 x TIFF: Run 1: 12sec (two passes), Run 2: 6sec (two passes)
50 x DNG: Run 1: 26sec (two passes), Run 2: 9sec (two passes)

Do these numbers seem right? I am not complaining! Just seems awfully fast. The final image looks fine - I presume it rendered it at 100% as I do not see that option any longer in the UI.

One more thing - I am sure you know this already and know that this is a BETA but on my 4k 32" screen in 1:1 mode, I can barely see the icons:
Screen Shot 2018-03-05 at 11.53.10.png
Screen Shot 2018-03-05 at 11.53.10.png (39.79 KiB) Viewed 727 times

User avatar
Stas Yatsenko
Posts: 3448
Joined: 06.05.2009 14:05
Contact:

Re: HF 25% slower on 10 core iMac Pro than 6 Core Mac Pro

Unread post by Stas Yatsenko » 06.03.2018 12:40

Yep, those times are the new reality. Seems reasonable for a 10-core machine.
You can speed it up even further by in-ticking "Show intermediary results" in the preferences.

And yes, thanks for pointing out the small icons, we're aware of it. It's work in progress.

Post Reply