Utilization of 2 or more GPU´s/GTX750ti #40027

New Issue

Rolf · 2014-05-04T23:03:02+02:00

Rolf commented

2014-05-04 23:03:02 +02:00

System Information
Windows 7 64bit
Intel I7-920 3,5ghz 16gb 2x GTX750ti

Blender Version
Broken: 2.70.xx, testet on fd80ac4
Worked: 2.69.11 eb4f2b4

Short description of error
Using more then one GPU, the utilization of both cards are bad.

In 2.69.11 eb4f2b4, i get speedup of ~90%-97%
In 2.70.5 fd80ac4 ~33%-60%
I think it has to do with this commit 1d016758

I made tests on Blenderartists, using mib2berlin´s Cornellbox 2.7x benchmark.
http://blenderartists.org/forum/showthread.php?327909-Cycles-NVidia-MAXWELL-Benchmarks&p=2639500&viewfull=1#post2639500

Exact steps for others to reproduce the error
Render with 2 GPU´s, look at the utilization of both cards.
www.blenderartists.org/forum/attachment.php?attachmentid=297482&d=1395468047

**System Information** Windows 7 64bit Intel I7-920 3,5ghz 16gb 2x GTX750ti **Blender Version** Broken: 2.70.xx, testet on fd80ac4 Worked: 2.69.11 eb4f2b4 **Short description of error** Using more then one GPU, the utilization of both cards are bad. ![gpu_load.jpg](https://archive.blender.org/developer/F86936/gpu_load.jpg) In 2.69.11 eb4f2b4, i get speedup of ~90%-97% In 2.70.5 fd80ac4 ~33%-60% I think it has to do with this commit 1d016758 I made tests on Blenderartists, using mib2berlin´s Cornellbox 2.7x benchmark. http://blenderartists.org/forum/showthread.php?327909-Cycles-NVidia-MAXWELL-Benchmarks&p=2639500&viewfull=1#post2639500 **Exact steps for others to reproduce the error** Render with 2 GPU´s, look at the utilization of both cards. www.blenderartists.org/forum/attachment.php?attachmentid=297482&d=1395468047

Rolf commented

2014-05-04 23:03:02 +02:00

Changed status to: 'Open'

Martijn Berger was assigned by Rolf

2014-05-04 23:03:02 +02:00

Rolf commented

2014-05-04 23:03:02 +02:00

Added subscriber: @RolfJawarsch

Thomas Dinges commented

2014-05-05 01:09:01 +02:00

Added subscriber: @ThomasDinges

Martijn Berger commented

2014-05-05 08:24:41 +02:00

Are both card dedicated to rendering or is one driving the display ?

We use different cuda command to try and keep the card as busy as possible depending on this.

How big is your tile size and how fast does one tile complete?

the asynchronous (2nd card and beyond) is depending on scheduling just enough work so the GPU comes back about every 1 second. the 1st card ( only if it drives the display) one CPU core is tasked to busy-waiting for the card to finish.

Are both card dedicated to rendering or is one driving the display ? We use different cuda command to try and keep the card as busy as possible depending on this. How big is your tile size and how fast does one tile complete? the asynchronous (2nd card and beyond) is depending on scheduling just enough work so the GPU comes back about every 1 second. the 1st card ( only if it drives the display) one CPU core is tasked to busy-waiting for the card to finish.

Rolf commented

2014-05-05 23:53:09 +02:00

Yes, one card is for driving the display.

Tilessize in this case was 256x256, but its also on 128/128 or 512x256 etc.
The greater the tilesize the worse the utilization.
The smaller the tilesize the longer the overall rendertime.

Each tile render different times by every render attempt with 2xGPU+busywait commit, it's hard to say, but i´ve tested it.
Mib´s Cornellbox for 2.7
2.70.5 fd80ac4

Rendertimes 2:30-2:42 4 Tiles@256x256 2xGPU
1.Top right: ~1:02 GPU1
2.Bottom right: ~1:19 GPU2
3.Top left: ~1:23 GPU1
4.Bottom left: ~1:22 GPU2

Rendertimes 3:44 4 Tiles@256x256 1xGPU
1.Top right: ~0:52
2.Bottom right: ~0:55
3.Top left: ~0:56
4.Bottom left: ~1:02

2.70.5 fd80ac4 without busywait commit
Rendertimes 1:58 4 Tiles@256x256 2xGPU
1.Top right: ~0:55 GPU1
2.Bottom right: ~0:56 GPU2
3.Top left: ~0:55 GPU1
4.Bottom left: ~1:03 GPU2

Rendertimes 3:52 4 Tiles@256x256 1xGPU
1.Top right: ~0:56
2.Bottom right: ~0:56
3.Top left: ~0:57
4.Bottom left: ~1:03

Yes, one card is for driving the display. Tilessize in this case was 256x256, but its also on 128/128 or 512x256 etc. The greater the tilesize the worse the utilization. The smaller the tilesize the longer the overall rendertime. Each tile render different times by every render attempt with 2xGPU+busywait commit, it's hard to say, but i´ve tested it. Mib´s Cornellbox for 2.7 2.70.5 fd80ac4 Rendertimes 2:30-2:42 4 Tiles@256x256 2xGPU 1.Top right: ~1:02 GPU1 2.Bottom right: ~1:19 GPU2 3.Top left: ~1:23 GPU1 4.Bottom left: ~1:22 GPU2 Rendertimes 3:44 4 Tiles@256x256 1xGPU 1.Top right: ~0:52 2.Bottom right: ~0:55 3.Top left: ~0:56 4.Bottom left: ~1:02 2.70.5 fd80ac4 without busywait commit Rendertimes 1:58 4 Tiles@256x256 2xGPU 1.Top right: ~0:55 GPU1 2.Bottom right: ~0:56 GPU2 3.Top left: ~0:55 GPU1 4.Bottom left: ~1:03 GPU2 Rendertimes 3:52 4 Tiles@256x256 1xGPU 1.Top right: ~0:56 2.Bottom right: ~0:56 3.Top left: ~0:57 4.Bottom left: ~1:03

Martijn Berger commented

2014-05-07 08:50:34 +02:00

Added subscriber: @brecht

Martijn Berger commented

2014-05-07 08:50:34 +02:00

I am not getting this behavior, thx for the detailed info though.

@ThomasDinges I think I need to talk with @brecht to see if we can further reduce the sync events and or if driving one GPU with sync api and another with async might be the cause.

I am not getting this behavior, thx for the detailed info though. @ThomasDinges I think I need to talk with @brecht to see if we can further reduce the sync events and or if driving one GPU with sync api and another with async might be the cause.

Thomas Beck commented

2014-05-09 16:59:08 +02:00

Added subscriber: @plasmasolutions

Thiébaut Champenier commented

2014-05-14 16:50:18 +02:00

Added subscriber: @Tibi-4

Thiébaut Champenier commented

2014-05-14 21:28:37 +02:00

Hi,

I’ve tested with two 750Ti cards not connected to any display, blender-2.70-30361a7-win64-vc12 and I get the following times for the cornell_bench_27:
One card 4’19"
Two cards 2’11"

Cheers!

Hi, I’ve tested with two 750Ti cards not connected to any display, blender-2.70-30361a7-win64-vc12 and I get the following times for the cornell_bench_27: One card 4’19" Two cards 2’11" Cheers!

Martijn Berger commented

2014-05-14 22:58:23 +02:00

It looks to me like mixing the synchronous and the asynchronous API has some additional downside.
This also implies that the current state is not really a working one.

It looks to me like mixing the synchronous and the asynchronous API has some additional downside. This also implies that the current state is not really a working one.

Brecht Van Lommel commented

2014-05-19 14:52:34 +02:00

Added subscriber: @MartijnBerger

Brecht Van Lommel commented

2014-05-19 14:52:34 +02:00

@MartijnBerger: do you think reverting 39bfde674c could solve the problem?

Not sure what else we can do here besides reverting the entire async system, I don't have the hardware to test this properly. And there's other complaints about updates taking too long, for which a more advanced system is probably needed.

@MartijnBerger: do you think reverting 39bfde674c could solve the problem? Not sure what else we can do here besides reverting the entire async system, I don't have the hardware to test this properly. And there's other complaints about updates taking too long, for which a more advanced system is probably needed.

Martijn Berger commented

2014-05-19 16:14:57 +02:00

@brecht. Yes I think we need a more advanced system. But current situation is less good then just doing busy waiting. Mixing both seems to have very undesirable effects and doing a proper fix requires more time and testing then I currently have.

I would just rip out the whole async handling for now. Ill re-add it to my long term todo.

@brecht. Yes I think we need a more advanced system. But current situation is less good then just doing busy waiting. Mixing both seems to have very undesirable effects and doing a proper fix requires more time and testing then I currently have. I would just rip out the whole async handling for now. Ill re-add it to my long term todo.

blender-admin commented

2014-05-19 19:33:09 +02:00

This issue was referenced by 3b53fffb77

This issue was referenced by 3b53fffb7788e19a0d05b8549aadbadf49279ca2

Brecht Van Lommel commented

2014-05-19 19:33:44 +02:00

Changed status from 'Open' to: 'Resolved'

Brecht Van Lommel closed this issue

2014-05-19 19:33:44 +02:00

Brecht Van Lommel commented

2014-05-19 19:33:44 +02:00

Closed by commit 3b53fffb77.

Closed by commit 3b53fffb77.

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Utilization of 2 or more GPU´s/GTX750ti #40027