Topic: Sample files for benchmarking?

[Ecuadorian]

Hi. in a few hours I'll be getting a new video card (GTX 570) so I'd like to know if there are sample files / benchmark files somewhere to do some comparison with my old card (9600 GT). I ask for sample files because I'm still testing the demo and thus I can't save anything.

 

If anyone could share a file, that would be good enough, but still, a specific scene for benchmarking purposes, like Octane Render has, would be best to compare GPUs and build a little database of the best and the worst for Lumion.

[Aaron]

There are quite a few recommendations on best GPU's flowing about already. To put it simple > The faster the gpu, the better the results.

Meanwhile, if you want to get yourself a rudimentary benchmark scene in the demo, just drop in 100 of the same speedtrees

[Ferry]

We will probably create some benchmark scenes at some point. For now you will have to creatng something yourself and look at the fps with effects on in theater mode.

[Remko]

Just Speedtree items would be a very biassed benchmark. Each shader has it's own characteristics. It's probably good testing speedtree but you should also test SSOA and many layers of the standard material. Another test would be a huge amount of geometry using relative simple shader (solid color for example)

[Ecuadorian]

Ok so I just bought a new EVGA card with the new NVIDIA GTX 570 GPU...

And I did a test before and after the upgrade:

 

With my old 9600 GT, Lumion takes 23.34 seconds to save each frame of a 1080p, maximum quality animation... of an empty scene with a river.

 

With my new GTX 570, which according to online sources should be almost equal in performance to the GTX 480, Lumion saves one animation frame each 4.36 seconds. Again, at 1080p, 3 stars quality, empty scene with river. No imported models, no trees, no camera movement, nothing at all, but it still takes 4.36 seconds per frame. I tested in Windows 7 64 bits, with a quad-core Core i5 750 CPU and 4GB of RAM.

 

I can't help but wonder what kind of futuristic alien technology you used to render the Lumion Nature with those gazillions of trees... in "1-2 seconds per frame".

[Ecuadorian]

I mean, the Lumion Nature video.

[Ferry]

Those 2 seconds where 2.5 seconds and they where at 720p. And it was the fastest alienware machine we could buy.

 

So try 720p to get a clearer comparison! I think it is probably similar to our insane machine.

[Ecuadorian]

Mmm… you're right. I take back my remark about "alien tech". Saving the same test as 720p takes only 2.18 seconds per frame.

 

I figured, then, that adding around a hundred objects, like trees, buildings, houses, and cats, might slow it down a lot… but no, even with all that stuff it's only taking 2.36 seconds per frame… Just an 8% difference.

 

I'm confused… Why is there so little difference in render time between an empty scene and a scene filled with stuff? Is the terrain system really so much of an overhead?

[Aaron]

Really also depends on what you call a lot of stuff. There seems to be a certain threshold that doesnt make too much of a difference in rendertimes per frame but the moment you start adding filters and high-poly geometry, there is a major impact most def (as is to be expected of course). 

I have some scenes containing many millions of polys, hundreds of trees, bushes etc. and when rendering those out including a lot of filters, the rendertime

for fullHD sometimes reaches 12seconds per frame. Which imo is still incredibly fast

[Ferry]

This is partly the because of realtime technology. We calculate per pixel all sorts of information always. That makes it a minimum speed which might seem slow. However adding content just changes the pixel not more pixels. Vertexes don't cause a big part of the render pipeline to be effected. Remember that 2.5 seconds is really long for a 3d card and it can push multi miliion amount of polygons with little effort in that time. Probably to get to 20 seconds per frame would only happen if you run out of memory. 

 

Also on your 9600gt the reason for the slow render time might have to do with running out of gfx memory.

 

So it is not that much the overhead more the what we do per pixel on your screen.

[Aaron]

How about drawcalls vs high poly objects Freddy?

 I've had it happen a couple of times that Lumion ran out of memory. Thats still a bit of a trial and error track for me.

 Not a complaint cause it was part of the test. 

[Ferry]

We are minimizing drawcalls where we can of course. If you are using 64 bit OS Lumion won't quickly say run out of memory. If you see that as a messagebox then that is not the 3d part but the cpu part.

[Ecuadorian]

You did an awesome job with the latest demo release!

 

I prepared a similar test (big mountain, lots of trees, a few houses) and now it takes only around 1.9 seconds per frame @1280x720.

[Ferry]

Thnx! We did fix a few performance issues besides allowing you to fix a number of problems in the first version. And new content for the full version of course.

[cpercer]

1.97 secs/frame

New scene -> "Big River" (Take 5 photos for movie, no camera movement, 3 star quality, 1280 x 720)

Lumion Update + RC1

PNY GTX580 "Enthusiast Edition" - No OC

Core i7 960 – No OC

12 gig RAM

Win 7 Pro x64 EN