Welcome, Guest
Username Password: Remember me

TOPIC: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 7 months ago #211108

  • hugly
  • OFFLINE
  • Platinum Boarder
  • Posts: 25426
  • 7 months ago
Hello Gr00by,

Thank you.

Have you exported the clip and the test sequence with 2020 on your current system, what does it say, compared to V14.5?
It's better to travel well than to arrive...

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211112

  • MarLondon
  • Pro User
  • OFFLINE
  • Expert Boarder
  • Posts: 120
  • 6 months, 4 weeks ago
Machine: Ryzen 7 1700X, 16 GB RAM, NVIDIA GTX 1060 3 GB, Lastest original NVIDIA Drivers

Kubutu 18.04:
Lightworks 14.5 test rendering time [1 min 52 sec]
Lightworks 2020.1 test rendering time [1 min]

Windows 10:
Lightworks 14.5 test rendering time [15 sec]
Lightworks 2020.1 test rendering time [11 sec]


Great improvement on Kubuntu with the new version, but still more than 5 times slower than Windows 10
{Rendered in more than 2 times the test video length)

Only 4 seconds improved on Windows 10
(Rendered in less than 1/2 of the test video length).

Similar to the @hugly Linux results on virual machine, strange, it's like the GPU is way slowed down.

This image is hidden for guests. Please log in or register to see it.

This image is hidden for guests. Please log in or register to see it.

This image is hidden for guests. Please log in or register to see it.

This image is hidden for guests. Please log in or register to see it.
Last Edit: 6 months, 4 weeks ago by MarLondon.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211129

  • G0bble
  • Pro User
  • OFFLINE
  • Gold Boarder
  • Posts: 292
  • 6 months, 4 weeks ago
Can you test with the project settings at 16bit? All the results posted earlier are at 16bit.

Thanks
G

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211130

  • MarLondon
  • Pro User
  • OFFLINE
  • Expert Boarder
  • Posts: 120
  • 6 months, 4 weeks ago
G0bble wrote:
Can you test with the project settings at 16bit? All the results posted earlier are at 16bit.

Thanks
G

The default setting was 8 bit, now I don't have time to reinstall previous versions and retest, I already wasted a lot of my time yesterday.
By the way I tested in 16 bit and the time was pratically the same on windows, and very few seconds of difference in linux, but pratically nothing compare to 1 minute of total rendering.
The real difference was using 32 bit floationg point, just in case.
Last Edit: 6 months, 4 weeks ago by MarLondon.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211131

  • MarLondon
  • Pro User
  • OFFLINE
  • Expert Boarder
  • Posts: 120
  • 6 months, 4 weeks ago
I overclocked a little bit CPU and RAM, same machine GPU NVIDIA GTX 1060 3GB

New results this time only using Lightworks 2020 beta:
(Same format H264 high HD 720p)


Windows 10:
[8 bit - 9 sec]
[16 bit - 10 sec]
[32 bit float p - 14 sec]

Kubuntu 18.04:
[8bit - 56 sec]
[16 bit - 1 min 08 sec]
[32 bit float p -1 min 08 sec] (the same of 16 bit no float p)

Don't ask me to reinstall Ltwk 14.5 to try because I don't want to try again, and for me is off topic for this topic too.
Last Edit: 6 months, 4 weeks ago by MarLondon.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211133

  • hugly
  • OFFLINE
  • Platinum Boarder
  • Posts: 25426
  • 6 months, 4 weeks ago
So far, it has been shown on different machines on Windows (MarLondon, Barus and myself) that exporting the test sequence with a GPU precision of 8-bit, 16-bit, and 16-bit float creates nearly the same export duration for all, whereas 32-bit float introduces a penalty of 50 percent.

I don't see this difference on Linux. Beside of other things, this can mean that 32-bit float simply isn't effective on Linux, but I don't know any procedure to verify this assumption.
It's better to travel well than to arrive...

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211134

  • MarLondon
  • Pro User
  • OFFLINE
  • Expert Boarder
  • Posts: 120
  • 6 months, 4 weeks ago
For me @hugly, there is a bottleneck in Linux on GPU data in/out, this is why 8, 16 bit and 32 bit float p values are so close.
Or something like that.
Probably only for Nvidia GPU.
Last Edit: 6 months, 4 weeks ago by MarLondon.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211136

  • hugly
  • OFFLINE
  • Platinum Boarder
  • Posts: 25426
  • 6 months, 4 weeks ago
I think that the data base collected so far doesn't allow any reliable conclusions about possible causes. We just know that there's a significant difference between Linux and Windows when exporting the test sequence on the same machine.

I tested Win10 1903 vs Fedor 30 both running on my system (AMD1920x 32GB, GTX1060 6GB) as virtual machines under VMware 15 with the same configuration. 8-Core CPU, 8GB of RAM. Export parameters: GPU precision 16-bit, 720p24 5Mbps.

Win10 Hardware clip/sequence: 6s/12s = 100fps/50fps = 50%

Win10 VMware clip/sequence: 9.5s/28s = 63fps/21fps = 34%

Fedora30 VMware clip/sequence: 11s/60s = 55fps/10fps = 18%

The difference between the virtual machines when exporting the sequence is appr. 50%.

By the way, it would be helpful to know the duration when exporting the clips, for both Linux and Windows under the same conditions, because that's an approximate measure of CPU performance.

Win10 Hardware:
This image is hidden for guests. Please log in or register to see it.

Win10 VMware:
This image is hidden for guests. Please log in or register to see it.

Fedora30 VMware:
This image is hidden for guests. Please log in or register to see it.
It's better to travel well than to arrive...
Last Edit: 6 months, 4 weeks ago by hugly.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211139

  • schrauber
  • OFFLINE
  • Platinum Boarder
  • Posts: 4395
  • 6 months, 4 weeks ago
MarLondon wrote:
..there is a bottleneck in Linux on GPU data in/out ..

In case you want to investigate further details:

hugly wrote:

- So you test with 10 times the same video, but in 10 different timelines.
I'm wondering if the data of 10 videos is transferred from the CPU to the GPU, or if only the data of one video is transferred, and the splitting of the 10 tracks is only GPU internal?
Does this differ between operating systems?
Maybe a second GPU test for comparison could bring further insights, in which only one video track is used, and the original 10 video tracks are simulated, in which one texture from one track is split into 10 paths in the routing?

- Test the influence of the shader performance: The current test contains many effects that produce only a minimal shader load, so that a large part of the GPU load is probably the GPU-internal data transfer between the many effects. If a powerful GPU still only has a low load in the test, then this probably doesn't matter.
Interesting could be test projects, mainly using sophisticated effects, where many calculations have to be done in the shader to see the influence of the shader performance.

- Test the influence of the sampler performance: For example, good blur effects take many samples of pixels. These samples require a relatively high GPU load, even if there are only a few calculations in the shader itself. It might be interesting to know if different GPUs on different operating systems have different sampler performances.
Mainly automatically translated
--------------------------------------------
Software: Lightworks 2020.1; || Windows 10, 64 Bit
Hardware: Intel i5-4440 (3,1 GHz); || shared RAM: 8 GB; || Intel HD Graphics 4600 (can use max. 2 GB of shared RAM)

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211141

  • gr00by
  • Pro User
  • OFFLINE
  • Platinum Boarder
  • Posts: 1151
  • 6 months, 4 weeks ago
Hi.

I think that the data base collected so far doesn't allow any reliable conclusions about possible causes.


That's why I think that further tests are worthless, until Lightworks devs diagnose the issue.

Regards,
Marcin
Canon C100 -- Manjaro Linux User

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211145

  • hugly
  • OFFLINE
  • Platinum Boarder
  • Posts: 25426
  • 6 months, 4 weeks ago
schrauber wrote:
- So you test with 10 times the same video, but in 10 different timelines.
I'm wondering if the data of 10 videos is transferred from the CPU to the GPU, or if only the data of one video is transferred, and the splitting of the 10 tracks is only GPU internal?

The tracks are shifted by some frames for a group of tracks. What you see are 3 different merged frames from the same source file plus the overlay from a second file

Maybe a second GPU test for comparison could bring further insights

Maybe, but with 2 people who can actively test on suitable hardware installations, G0bble and MarLondon, currently, one Pro one Free license, personally I think it makes no sense to change the test conditions.
It's better to travel well than to arrive...
Last Edit: 6 months, 4 weeks ago by hugly.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211146

  • hugly
  • OFFLINE
  • Platinum Boarder
  • Posts: 25426
  • 6 months, 4 weeks ago
gr00by wrote:
That's why I think that further tests are worthless, until Lightworks devs diagnose the issue.

I think what we collect here is just a heads up for devs that there are people who care about Linux performance and I strongly believe that the Editshare developers are aware of the issue. I'm not sure if anybody on this planet is able to describe the causes, even with hands on the code.

Edit: And, to be able to describe possible causes, does by no means imply that there's a fix.
It's better to travel well than to arrive...
Last Edit: 6 months, 4 weeks ago by hugly.

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211151

  • hugly
  • OFFLINE
  • Platinum Boarder
  • Posts: 25426
  • 6 months, 4 weeks ago
On question to the Linux community: Is there any NLE on Linux which performs better than Lightworks 2020?
It's better to travel well than to arrive...

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211152

  • gr00by
  • Pro User
  • OFFLINE
  • Platinum Boarder
  • Posts: 1151
  • 6 months, 4 weeks ago
hugly wrote:
On question to the Linux community: Is there any NLE on Linux which performs better than Lightworks 2020?


No.
Canon C100 -- Manjaro Linux User

Re: Huge GPU tests difference between Linux and Windows Lightworks 14 on the same machine 6 months, 4 weeks ago #211159

  • MarLondon
  • Pro User
  • OFFLINE
  • Expert Boarder
  • Posts: 120
  • 6 months, 4 weeks ago
hugly wrote:
On question to the Linux community: Is there any NLE on Linux which performs better than Lightworks 2020?

There is DV Resolve, but I never compared directly the performance with Lightworks on Linux.
So I don't know.

Edit: By the way, for my opinion, if you want to use Lightworks is better use it with Windows, unless you have no choice.
But when you can use Windows, the choice of NLE could become bigger.
Last Edit: 6 months, 4 weeks ago by MarLondon.
Time to create page: 0.46 seconds
Scroll To Top