Intermediate multi-gpu test version for windows x64

[radiance]

Hi all,

*** THIS IS NOT A RELEASE ***

This is a little test binary for windows x64 only,
and it's not recommended for use for any use other than seeing it works with multi-gpu systems,
by people interested in tinkering with things.


We would'nt recommend to use this binary for any commercial work as it is not fully finished/tested.


This is a version of octane that can use more than one GPU at a time during rendering.
However there are a few limitations still that we will iron during the days ahead.

If you don't have multiple GPU's in your system there is no need to play with this version.

The difference between this version and the previous 2.1 release is that you can add and remove devices to the 'active' device list,
in the updated device manager.

When rendering with multiple GPUs, the rendering runs in parallel.

We have not parallellized the tonemapping kernels yet, so it's not yet fully optimized and it's not a 2x speed increase with 2 GPUs.

Also, major areas of the engine have been rewritten and not all functionalities are hooked up and tested yet,
so some things like material picking etc might not work as they should yet.

A few notes:

- the total available memory for use in octane is the minimum of your cards,
so if you have for example 2 cards with 1GB ram, and one card with 256 MB ram, and you add all 3 for rendering, you will only be able to use 256 MB ram.
In this scenario we'd recommend you use the 2 cards together and leave out the 3rd one.

- Currently the load is not adjusted per GPU power, so if you have 2 fast cards and one slow card, they will all 3 get the same amount of work to do,
and your slow card will make the other cards wait a bit every frame. we're improving this so that the load is propperly balanced in the next versions.

This version will work best if you're cards are all of the same type.

- rendering is parallelized, but not tonemapping, so on large resolutions the speedup of multiple cards will become lower,
we are already in the process of fixing this, so octane will effectively have a linear speedup when adding more cards. (or near-linear)

- the navigation/camera manipulation is not %100, it's a bit jerky, we're fixing this.

You need to make sure SLI is turned off to get octane to see multiple GPUs

Also, Please don't use this thread to report other unrelated problems,
only issues specific to getting this release to see your multiple GPU setup, and statistics.
There's no point in reporting issues of 2.1 that we are already aware of.

If you want to post statistics, please run the standardized trench benchmand scene, with the OCS project settings (no changes, eg pathtracing)
Also, please do not use this binary for any off site statistics or documentation, as the speedup/performance is not finalized and will improve.
What we'd like to hear from people trying it is if it works, if it can see all the GPUs, if it's stable, and some benchmark results

Thanks,
Radiance

[mlody47]

I love You

[mlody47]

"and speed up will improve more??? hell this is already fast as hell""


when i was typing this sentence "and speed up will improve more??? hell this is already fast as hell""

I got finished render already

[radiance]

"and speed up will improve more??? hell this is already fast as hell""


when i was typing this sentence "and speed up will improve more??? hell this is already fast as hell""

I got finished render already

can you post a screenshot of the benchmark scene rendering in progress so we can see the megasamples/px ?
would be handy if you can make 2, one with 1 GPU and one 2 the 2nd GPU.

Radiance

[mlody47]

But You know one thing is weird and great . Im rendering on 2 gpus and still no lag on ui or anything. maybe barely noticeable

[mlody47]

ok some runtime error come up when opening new scene when rendering image.

[radiance]

ok some runtime error come up when opening new scene when rendering image.

yeah, we know this...

i'd be interested to see some benchmarks, eg one render with 1 gpu and one with 2 gpus.
using the standardized benchmark trench scene in the demosuite, while rendering, so we can see the megasamples/sec.

also a shot of the device manager

you need to load the benchmark scene, then change the device setup (eg add 2nd gpu to used list),
then start the engine by clicking on the mesh node...

Radiance

[andrian]

There is something weird, or I do something wrong, but seems I get higher samples with first single GPU than with both...
And what about memory usage?? It's double in multi GPU mode. I know that have to be loaded in both GPU-s.. but that way seems i have left less memory than with one GPU scenario. Also some weird line in the middle, separating both GPU calculations in multi GPU mode. Renders was done with pathtracing kernel, scene is saved in Octane beta 2.1 non multi GPU version. No material applied to any object ( uses the kilaD scipt to export ).

multi GPU FPS-Samples.jpg (8.79 KiB) Viewed 3977 times

first GPU FPS-Samples.jpg (11.53 KiB) Viewed 3977 times

secong GPU FPS-Samples.jpg (11.46 KiB) Viewed 3977 times

[acc24ex]

Using the benchscene:
I got
1.8 Msamples/sec on 260 gtx
0.83 on 8800 gtx
and 1.12 on 8800gtx + 260 gtx

Not doing much on my system as well.


someone with two 480 cards should do this test..

But it runs ok sort off, managed to load some scenes, didn't crash .. much

[YoonKyung]

1x GTX480 = 4.05 msps : +4.05(100%), 1.00(100.0%)
2x GTX480 = 7.80 msps : +3.75(92.5%), 1.93(96.50%)
3x GTX480 = 11.20 msps : +3.40(83.9%), 2.77(92.22%)

Near perfect scalability !!

The GPU scalability was near linear, perfect.
It was really fast. Scenes were completed within 10 seconds.
And GPU selecting dialog box was working well.

But I found that there is not the 'File Close' menu.
And when I change the resolution or do 'File New' command, octane do abnormal exit.
And [File][Exit] takes long time.