PDA

View Full Version : New Benchmark Scene - Superballs



Scazzino
02-27-2012, 02:29 PM
I needed a good benchmarking scene that I could use to test the render speeds of my various Mac systems under heavy render loads and I also needed a simple render test scene for my DLI_SNUB-Launcher tutorial/white paper that I'm updating for LW11. So I created this superballs scene. Once I finish updating my tutorials I'll release the scene itself on the DLI_SNUB-Launcher download page too.

It uses lots of features that really stress the cores/threads to the max including radiosity, blurred reflections, blurred refractions, SSS, DOF and 600 instances that exercise all those features (though you only see a small subset of all the instances in this particular image there are many more visible in all the reflections).

This took 3 Hrs, 54 Min, 9 Secs to render on the 3GHz 8-core Mac Pro
and 1 Hr, 34 Min, 40 Secs to render on the 2.93 GHz 12-core Mac Pro (while I was also using the 12 core to do other things. ;-)

http://dreamlight.com/media/graphics/3d/images/DLI_SuperBalls-Comp-1280.jpg

Scazzino
02-27-2012, 02:32 PM
BTW: This was also a way for me to test LW11's new instancing, which is pretty cool! The yellow ball near the center is the only real ball and the remaining 600 are new LW11 instances. :lightwave

cagey5
02-27-2012, 02:41 PM
That's a very nice image. :thumbsup:

i dread to think how long it would take my computer to render that.

Waves of light
02-27-2012, 03:27 PM
I'm quite shocked with the time save between an 8 core and 12 core set up.

My rig would wake me up in the morning to tell it was done. Let us know when the scene is available and I will stress the poor girl out. Just make sure it's in the next 26 days, coz that's all I got to play with LW11.

Did you use the nodel set up in the manual to change the colours using the 'multi switch' node?

JonW
02-27-2012, 04:26 PM
A x55xx/x56xx CPU is a lot quicker than a x54xx CPU or even more so than a x53xx CPU. The new E5-x26xx CPUs look like they are going to be a lot quicker than the 55/56. A Ghz gets more & more efficient with each generation.

A new dual E5-2687w 16/32 core box is going to scream!



It would be nice if there was a scene that could be a "sticky" for benchmarking it's not a good look everyone using Cinebench!

JonW
02-27-2012, 04:34 PM
This took 3 Hrs, 54 Min, 9 Secs to render on the 3GHz 8-core Mac Pro
and 1 Hr, 34 Min, 40 Secs to render on the 2.93 GHz 12-core Mac Pro (while I was also using the 12 core to do other things. ;-)



I tried the Grass & LW Logo scene on my W5580 box (25.6 GHz total) 47 seconds. So this is proportional to a 2.93 X5570 12 core box.

Scazzino
02-27-2012, 07:53 PM
Thanks guys! :thumbsup:

@ Cagey: I wanted a scene that would use a bunch of relatively compute expensive features to get a good comparison of the various machines with various numbers of cores and threads.

@Waves: I used the first method of using the color gradient (with stepped keys) driven from instance info: Fixed Random to change the color of all the balls. I've posted an image of the part of the node map that changes the colors.

@JonW: Yes I can't wait to see if Apple does release a new 16-core/32-thread beast soon. :D

Scazzino
02-29-2012, 01:12 PM
I was having too much fun with this image that I couldn't resist tweaking it a bit more. This one uses SSS, a higher refraction angle and two bounce radiosity rather than Kappa and one bounce.

Which do you all prefer, this one or the previous version? :question:

http://dreamlight.com/media/graphics/3d/images/DLI_SuperBalls-Comp-1280-B.jpg

Waves of light
02-29-2012, 01:58 PM
Thanks for the nodel set up. I must admit I prefer the first image as it looks more like bouncing balls. I think the second is creeping slowly to glass marbles. IMO, lol.

Scazzino
02-29-2012, 02:06 PM
Good point! :thumbsup:

Yes the real balls fall somewhere between the two. I have a bunch of them here. They are more translucent than the first image and not quite as much as the second image. The tricky part is controlling the internal lighting that's refracting from the lower portion on the inside. In the real balls it's more of a soft diffuse glow. In the renders it gets blown out a bit from the HDR window highlights from the HDR image probe. The tough part is that tweaking one thing, usually affects something else, making it tough to get it all working the way you want. I am post-processing the resulting HDR linear image in Shake a little bit to compress the overall tonal range to bring the HDR highlights back into normal color space. One thing that I wanted to get working better was the glow under the balls. The real balls have a nice soft glow beneath them from the light that refracts through them but I had a tough time trying to replicate that effect since no matter what I did those transmitted hotspots were cast further away from the balls than in reality. I tried various refraction amounts from 1-3 without much luck.

bazsa73
02-29-2012, 10:28 PM
I like both, the first is plastic the second is made of harder material and the lighting is stronger which I appreciate.
In the morning lighting could be just like that. Nice renders.:thumbsup:

Scazzino
03-02-2012, 12:53 PM
Thanks! :thumbsup:

And because I guess I'm never quite satisfied with my own work... here's version C. ;)

I just had to try a few more tweaks. This one uses a new linoleum tile floor, tweaked settings, a true Z-buffer to boost the in-camera DOF (rather than a simple gradient) and a slight layer of dust. Everything is created from procedural nodal textures other than the HDR kitchen probe and a photo of a linoleum tile that I took in my own kitchen.

So which is everyone's favorite, A, B or C? And why?

http://dreamlight.com/media/graphics/3d/images/DLI_SuperBalls-Comp-1280-C.jpg

JonW
03-02-2012, 03:32 PM
Colour balance, tiles & balls are great on the last image. I feel that the balls in the distance, the lighting/glow whatever it is, doesn't look quite right. Somehow it needs to be cut back a bit.

Never the less it's extremely impressive!

Scazzino
03-02-2012, 07:38 PM
Thanks Jon! :thumbsup:

Yes, that bothered me too. Believe it or not, that is taken down a bit already. I have the reflections being diminished based on distance. I could probably increase that and also tone down the transparency. The problem arises because as the angle flattens in the distance, the balls are refracting more of the highlights from the HDR Light Probe and blowing out a bit too much.

Scazzino
03-02-2012, 07:51 PM
Actually I could also diminish the refraction index a bit based on distance, that would also cut down on it blowing out too much in the distance as the viewing angle flattens out. I also used a slight fog effect to darken the balls in the distance a bit too.

First I might try toning down the distant highlights a bit more in post using the depth buffer.

Scazzino
03-02-2012, 08:58 PM
OK, I've toned down the highlights a bit more in the distance using the depth buffer in Shake and updated the third image above. :thumbsup:

Waves of light
03-03-2012, 09:43 AM
Still no.1 for me. IMO they are the ones that look more like bouncing balls, because of the diffuse settings and less transparency.

Scazzino
03-05-2012, 10:27 AM
Here are the render times for the third version.

8-core 3GHz Mac Pro (first version of the 8-core, 2007): 6h 19m 43s (22783.9 seconds)
12-core 2.93GHz Mac Pro: 2h 9m 53s (7793.5 seconds)

It's currently rendering on my 4-core 2.2 GHz 15" MacBook Pro and I'll post the results when that's done. Looks like it'll be similar to the 8-core Mac Pro.

Scazzino
03-06-2012, 11:25 AM
4-core 2.2GHz i7 15" MacBook Pro (mid-2011) took 6h 34m 55s (23695.9 seconds) on the third image.

I'm currently working on a beta version of the DLI_SNUB-Launcher update. When it's ready I'll put this scene together with the beta download for testing. Then you can run the same render on your machines to compare render speeds. :beerchug:

Andrewstopheles
03-08-2012, 06:07 AM
The third, C, is my favorite. They look like "jet-balls" from a vending machine.

bazsa73
03-08-2012, 07:19 AM
Non-noisy DOF is still only achievable by setting AA to the extreme?

Scazzino
03-08-2012, 10:59 AM
Pretty much. One trick I use to minimize it though is to render with a low to moderate amount of in-camera DOF at a modest to low AA setting but also save out a depth buffer. I then use that depth buffer in post-processing in Shake to apply a further depth based blur to help smooth out the graininess of the DOF effects and to push it even further if desired. That way you get true DOF so that foregrounds/backgrounds blur properly, but you also get the ability for some real-time adjustability in post-processing. :thumbsup:

Scazzino
03-08-2012, 11:23 AM
This two-step DOF approach works well as a good compromise between accurate in-camera DOF and faster post-processing DOF. The problem with using just post-processing DOF is that the edges of foreground objects don't blur correctly against distant backgrounds because you don't have the data of what the pixels behind the object edge would look like (unless you are also rendering in multiple foreground/background passes). Using this two-step approach gives accurate edge blurring though a bit noisy unless you have very high AA. Then the additional post-process blur helps smooth out that noise and pushing the DOF effect further in post is OK because you have some of the data that you'd need from behind those foreground object edges in the image. For the fastest, best post-process DOF you can of course render the foreground/background in multiple passes, but this can get quite tricky if you have lots of overlapping objects at various distances and the added complexity of the scene setups can negate the benefit of the faster rendering. :hey:

Scazzino
03-21-2012, 10:54 PM
OK. I've posted the new DLI_SuperBalls Benchmark scene along with a beta version of DLI_SNUB-Launcher at the following URL for download:

Direct Download Link (http://dreamlight.com/downloads/DLI_SNUB-Launcher%201.8beta.zip)

DLI_SNUB-Launcher Page (http://dreamlight.com/webshop/utilities/snublauncher.html)

Enjoy! :thumbsup:

Scazzino
03-21-2012, 11:20 PM
Feel free to post your render times for the DLI_SuperBalls benchmark scene here along with your system specs. :beerchug:

3dworks
03-22-2012, 04:40 AM
i followed your instructions but it seems not to work properly here, lwsn runs fast through the scene and reports 1.2 sec render time...

see log here


Last login: Thu Mar 22 11:33:44 on ttys000
lion:~ 3dworks$ "/Production/Lightwave/LW3D_11_0/bin/lwsn" -3 -c"/Production/Lightwave/LW3D_11_0/Preferences" -d"/Volumes/Lion_W/Projects_Lion/Test_Lion/lightwave/DLI_SuperBalls/" "/Volumes/Lion_W/Projects_Lion/Test_Lion/lightwave/DLI_SuperBalls/Scenes/DLI_LW11-SuperBalls_01H.lws" 1 1 1
LightWave MachOMac ScreamerNet Module (Build 2238)
CPU number: 669

Current directory is now "/Volumes/Lion_W/Projects_Lion/Test_Lion/lightwave/DLI_SuperBalls/".
Loading "/Volumes/Lion_W/Projects_Lion/Test_Lion/lightwave/DLI_SuperBalls/Scenes/DLI_LW11-SuperBalls_01H.lws".
Clearing scene
Loading scene
Pre-processing scene
Loading plug-in
Plug-in class "MasterHandler" name ".SceneEditorStandardBanks" slot 1
Plug-in loaded
Loading plug-in
Plug-in class "MasterHandler" name "SceneEditor" slot 2
Plug-in loaded
Loading plug-in
Plug-in class "ViewportObjectHandler" name ".VPR" slot 1
Loading plug-in
Plug-in class "ViewportObjectHandler" name ".VPR" slot 2
Loading plug-in
Plug-in class "ViewportObjectHandler" name ".VPR" slot 3
Loading plug-in
Plug-in class "ViewportObjectHandler" name ".VPR" slot 4
Loading "Objects/DLI_Floor_01F.lwo"
Loading plug-in
Plug-in class "InstancerHandler" name "InstanceGenerator" slot 1
Plug-in loaded
Loading "Objects/DLI_SuperBall_01G.lwo"
Loading "Objects/DLI_Floor_01F.lwo"
Loading plug-in
Plug-in class "InstancerHandler" name "InstanceGenerator" slot 1
Plug-in loaded
Loading plug-in
Plug-in class "LightHandler" name "SpotLight" slot 1
Plug-in loaded
Loading plug-in
Plug-in class "CameraHandler" name "Perspective" slot 1
Plug-in loaded
Loading plug-in
Plug-in class "EnvironmentHandler" name "LW_ImageWorld" slot 1
Plug-in loaded
Loading plug-in
Plug-in class "VirtualStudio" name "VirtualStudio" slot 1
Plug-in loaded
Validating scene
Preparing "Objects/DLI_Floor_01F.lwo"
Preparing "Objects/DLI_SuperBall_01G.lwo"
Preparing "Objects/DLI_Floor_01F.lwo"
Scene loaded.
Allocating frame buffers.
Allocating segment buffers.
Frame: 1.
Segment: 1/1.
Pass: 1/1.
Updating geometry.
Moving DLI_Floor_01F (1).
Moving DLI_SuperBall_01G.
Moving DLI_Floor_01F (2).
Preprocessing Frame....
Computing preliminary radiosity solution....
Rendering frame 1, segment 1/1, pass 1/1.
Image Reconstruction.
Image Reconstruction : 0%.
Image Reconstruction : 1%.
Image Reconstruction : 2%.
Image Reconstruction : 3%.
Image Reconstruction : 4%.
Image Reconstruction : 5%.
Image Reconstruction : 6%.
Image Reconstruction : 7%.
Image Reconstruction : 8%.
Image Reconstruction : 10%.
Image Reconstruction : 11%.
Image Reconstruction : 12%.
Image Reconstruction : 13%.
Image Reconstruction : 14%.
Image Reconstruction : 15%.
Image Reconstruction : 16%.
Image Reconstruction : 17%.
Image Reconstruction : 18%.
Image Reconstruction : 20%.
Image Reconstruction : 21%.
Image Reconstruction : 22%.
Image Reconstruction : 23%.
Image Reconstruction : 24%.
Image Reconstruction : 25%.
Image Reconstruction : 26%.
Image Reconstruction : 27%.
Image Reconstruction : 28%.
Image Reconstruction : 30%.
Image Reconstruction : 31%.
Image Reconstruction : 32%.
Image Reconstruction : 33%.
Image Reconstruction : 34%.
Image Reconstruction : 35%.
Image Reconstruction : 36%.
Image Reconstruction : 37%.
Image Reconstruction : 38%.
Image Reconstruction : 40%.
Image Reconstruction : 41%.
Image Reconstruction : 42%.
Image Reconstruction : 43%.
Image Reconstruction : 44%.
Image Reconstruction : 45%.
Image Reconstruction : 46%.
Image Reconstruction : 47%.
Image Reconstruction : 48%.
Image Reconstruction : 50%.
Image Reconstruction : 51%.
Image Reconstruction : 52%.
Image Reconstruction : 53%.
Image Reconstruction : 54%.
Image Reconstruction : 55%.
Image Reconstruction : 56%.
Image Reconstruction : 57%.
Image Reconstruction : 58%.
Image Reconstruction : 60%.
Image Reconstruction : 61%.
Image Reconstruction : 62%.
Image Reconstruction : 63%.
Image Reconstruction : 64%.
Image Reconstruction : 65%.
Image Reconstruction : 66%.
Image Reconstruction : 67%.
Image Reconstruction : 68%.
Image Reconstruction : 70%.
Image Reconstruction : 71%.
Image Reconstruction : 72%.
Image Reconstruction : 73%.
Image Reconstruction : 74%.
Image Reconstruction : 75%.
Image Reconstruction : 76%.
Image Reconstruction : 77%.
Image Reconstruction : 78%.
Image Reconstruction : 80%.
Image Reconstruction : 81%.
Image Reconstruction : 82%.
Image Reconstruction : 83%.
Image Reconstruction : 84%.
Image Reconstruction : 85%.
Image Reconstruction : 86%.
Image Reconstruction : 87%.
Image Reconstruction : 89%.
Image Reconstruction : 90%.
Image Reconstruction : 91%.
Image Reconstruction : 92%.
Image Reconstruction : 93%.
Image Reconstruction : 94%.
Image Reconstruction : 95%.
Image Reconstruction : 96%.
Image Reconstruction : 97%.
Image Reconstruction : 99%.
Writing RGB image to /Volumes/Lion_W/Projects_Lion/Test_Lion/lightwave/DLI_SuperBalls/Renders/DLI_LW11-SuperBalls_01H_001.exr.
Frame completed.
Last Frame Rendered: 1.
Rendering Time: 1.2 seconds.
Freeing segment buffers.
Freeing frame buffers.
lion:~ 3dworks$

3dworks
03-22-2012, 04:44 AM
btw nice scene, but imo it's taking too much time for becoming a popular benchmark. nobody will easily dedicate more than 2 h on even the fastest machines around for just checking speed. something around 5-10 min is the maximum i would dedicate.

Scazzino
03-22-2012, 08:58 AM
Thanks 3D Works. That's a good point for quick testing.

I specifically made this scene for testing high end Macs using heavy duty features exactly because most benchmark scenes take way too little time. Especially as the Macs get faster and faster. This scene takes about 2 hours on the top line 12 core Mac Pro. As the Mac's get faster and faster that time will dramatically drop (16 core/32 threads is rumored in the works), so I wanted a scene that would be useful for a while. The problem with short benchmarks is that the shorter the time gets the less accurately they scale when trying to compare really fast machines. This one is intended to run overnight, much like a high resolution complex 3D illustration might be run.

You can easily drop the AA and/or resolution though to dramatically speed up the renders on a slower machine or for a quick test though. :thumbsup:

Thanks for testing DLI_SNUB-Launcher too. I'm not seeing that problem here. I'll look into it and let you know what I find. Have you used lwsn successfully on your current Mac/configuration without using DLI_SNUB-Launcher (just to be sure that lwsn and your plugins etc are installed and working properly)?

Scazzino
03-22-2012, 09:01 AM
What did the rendered frame look like? All black? Even though the output said it loaded the scene and objects if it "rendered" instantly then it's possible that nothing really loaded to render.

Scazzino
03-22-2012, 09:13 AM
Oh, one thing I could do however is set the resolution/AA lower and get those times on my Macs. Then when I publish it I can include both scenes, the original High Res and a newer Low Res version. That would help satisfy the entire range of benchmarking from a quick test to a torturous workout.

Thanks for the feedback, keep em coming! :thumbsup:

Scazzino
03-22-2012, 10:22 AM
BTW: That lwsn issue could be related to Lion. I'm still running on Snow Leopard, though I do have Lion installed on the 12-core. I'll run some tests there and see what I find.

Scazzino
03-22-2012, 11:50 AM
Hmm. The lwsn issue doesn't look like a Lion issue. I just ran the test on Lion and everything went fine.

BTW: I dropped the resolution to 25% and the AA to 8 and that dropped the time to about 4 minutes on the 12-core. Though I may adjust the scene for non-linear/non-HDR colorspace for the low res test version so that the raw render will look better without bringing it into an HDR compositor like Shake.

Scazzino
03-22-2012, 02:44 PM
OK: I've updated the DLI_SNUB-Launcher 1.8beta downloader with a new set of benchmark scenes.

A LowResolution version that is set to 25% with lower AA, lower radiosity and no DOF. It's also set to render to a simple JPG with limited dynamic range. That only takes 4.5 minutes on the 12-Core Mac Pro.

And the original High Resolution version. That one is set to 100% with all the bells and whistles. That one renders to an HDR EXR with an optional Depth Buffer intended for post-processing. It takes about 2 hours to render on the 12-Core Mac Pro.

Direct Download Link (http://dreamlight.com/downloads/DLI_SNUB-Launcher%201.8beta.zip)

DLI_SNUB-Launcher Page (http://dreamlight.com/webshop/utilities/snublauncher.html)

:thumbsup:

Scazzino
03-22-2012, 02:54 PM
Feel free to post render times for the LowRes and/or HighRes scene, just be sure to include your system specs and which version you're rendering. I'd be curious to see the times across a broad range of systems, both Mac and Win. :thumbsup:

Here are times for both the Low and High Res benchmarks on a few of my Macs:

DLI_LW11-SuperBalls-LowRes.lws

12-core 2.93GHz Mac Pro: (MacPro5,1: Mid 2010, Dual Six-core Intel Xeon X5670 Gulftown)
4m 36s (276.4 seconds)

8-core 3GHz Mac Pro (MacPro2,1: Mid 2007, Dual Quad-core Intel Xeon X5365 Clovertown)
10m 9s (609.8 seconds)

4-core 2.2GHz 15" MacBook Pro (MacBookPro8,2: Mid 2011, 4-core Intel Core i7 2720QM Sandy Bridge)
13m 6s (786.0 seconds)

DLI_LW11-SuperBalls-HighRes.lws

12-core 2.93GHz Mac Pro: (MacPro5,1: Mid 2010, 2x 6-core Intel Xeon X5670 Gulftown)
2h 9m 53s (7793.5 seconds)

8-core 3GHz Mac Pro (MacPro2,1: Mid 2007, 2x 4-core Intel Xeon X5365 Clovertown)
6h 19m 43s (22783.9 seconds)

4-core 2.2GHz 15" MacBook Pro (MacBookPro8,2: Mid 2011, 4-core Intel Core i7 2720QM Sandy Bridge)
6h 34m 55s (23695.9 seconds)

JonW
03-22-2012, 10:59 PM
2 x W5580 3.2GHz XP64 2h 58m 51s (10731.3 seconds)

Did the same scene with AA16 3h 19m 57s (11997.6 seconds)
A few bits of the floor looked different but both looked fine.

Markc
03-23-2012, 04:23 PM
In LW10.1, is there any way to use all my cores in standalone mode?
I am on a Quad MacPro (8 threads), but only seem to be using about 35% cpu to render, also should the app not be running in 64bit?

Scazzino
03-23-2012, 04:49 PM
Yes 10.1 should run in 64 bit mode, on Mac OS X Snow Leopard or later. Make sure "Open in 32 bit mode" is not checked in the get-info.

Then set Multithreading in the Render Globals Render tab to automatic and it should render on all cores/threads. Or you can specifically set it to a specific number of threads if you like.

Markc
03-23-2012, 05:16 PM
Thanks Mike.
I tried a different scene file, and it worked perfectly using full cpu threads. :thumbsup:

I also realised it was Lwsn doing the rendering (in 64bit) not the snub launcher application......doohh.

Scazzino
03-23-2012, 08:29 PM
Right DLI_SNUB-Launcher only builds the command line and launches lwsn. There's no need for DLI_SNUB-Launcher itself to be 64 bit, lwsn will launch in 64 bit and render the scene using all available threads if multithreaded is set to automatic (which it is by default unless you change it in Layout).