How to get the MP advantage

NewTek Forum: LightWave 3D®: Mac LW: How to get the MP advantage
Top of pagePrevious messageNext messageBottom of pageLink to this message   By Burner (65.7.229.79) on Sunday, June 03, 2001 - 06:44 pm:

I just bought a G4 DP 533 machine and I'm trying to figure out how to gain any advantage using multi-threading.

So far I don't see any advantage at all to using an MP Mac. I'm new to LW, so maybe I'm missing something simple.

Here's a link to a page detailing my results. You can also download my test file and give it a whirl yourself.

http://www.bennettinnovations.com/users/cweyers/r_tests/

Top of pagePrevious messageNext messageBottom of pageLink to this message   By Daniel (158.252.223.56) on Sunday, June 03, 2001 - 10:04 pm:

The scene file is pretty simple.

Add Hypervoxels and some area lights. Add more geometry to. anywhere from 50,000 to 200,000 polys or more.

You might see the difference then.

Dan

Top of pagePrevious messageNext messageBottom of pageLink to this message   By shawn (24.20.51.213) on Sunday, June 03, 2001 - 10:53 pm:

I agree you will see no gain on simple scenes. Mp is only good with raytraceing and more complex stuff. If you add some reflections and refraction
you should see as much as 60 percent decrease in rendering times from one thread. From what I understand their is some overhead for mp so simple
stuff can actualy be slower in some cases. I have a
duel 500 and have seen great gains from multiple threads.

shawn ogle

Top of pagePrevious messageNext messageBottom of pageLink to this message   By burner (216.37.4.71) on Monday, June 04, 2001 - 08:40 am:

MP should be an advantage regardless of scene complexity.

Where I was really hoping MP would benefit me in LW was in anti-aliasing passes. They seem very slow in LW compared to other applications.

I tested the same scene in LW on my g3 500 powerbook. It took 6.8 seconds.

I also tested the same scene in the demo version of Cinema 4DXL7. On the powerbook the render time in XL was 9 seconds. On the DP 533 machine the render time was 5 seconds. I could watch in picture viewer as the render was divided in half.

I also did a quick test in Maya (not with this scene). I've got a Dual 450 PIII at work. Rendering a simple sphere casting a soft shadow on a plane with high anti-aliasing took 4 seconds with one processor and 2 seconds with both processors.

Top of pagePrevious messageNext messageBottom of pageLink to this message   By Teksoup (207.235.86.2) on Monday, June 04, 2001 - 08:59 am:

Multiprocessing has everything to do with scene complexity. The overhead for using a dual processors on a simple scene to enough to notice little difference. Any good test would include a scene with some complexity. I don't know what scene you were using to test or how fast it rendered in LightWave but as a general rule any frame that is going to render in under 10 second you are not going to see an advantage. Try your test using the blade scene in the space folder of the content you will notice a significant difference. If it is anti-aliasing that you want to test adjust the anti-aliasing in that scene you should see a difference.

Top of pagePrevious messageNext messageBottom of pageLink to this message   By david lubell (Dlubell) (12.96.176.35) on Monday, June 04, 2001 - 09:10 am:

in order to see a speed decrease in processing times you must change 1 thread renders to 2 thread in the render panel..try it on any scene you should see a significant difference between 1 and 2 thread renders..
good luck
dl

Top of pagePrevious messageNext messageBottom of pageLink to this message   By Burner (216.37.4.71) on Monday, June 04, 2001 - 09:47 am:

I added a volume light to the scene and tried my tests again. As stated there was a much bigger difference.

1 thread = 320 seconds
8 threads= 219 seconds

(There were varying speed increases with 2 and 4, but 8 was fastest)

I still don't understand why multi-processing isn't helping out AA passes. If I'm just rendering a simple logo or type piece I should be able to take advantage of my MP machine just as much as if I were rendering a more complex scene.

On adjusting anti-aliasing: Is there somewhere to adjust AA so it can take better advantage of multi-processing? What I saw in my tests is that using multiple threades doesn't impact the speed of AA passes at all. I could bump the AA up to "high" and adding more threads didn't help.

Top of pagePrevious messageNext messageBottom of pageLink to this message   By Burner (216.37.4.71) on Monday, June 04, 2001 - 10:45 am:

One more test.

I took the same scene and rotated the panel 360 over 15 frames and turned on normal motion blur set to it's defaults.

1thread = 28.1 seconds
2threads= 25.9 seconds
4threads= 25.1 seconds
8threads= 25.2 seconds

It seems like there's very little happening with MP and pixel filters. I would have expected a much speedier render under multiple threads.

Also it appears there's no rhyme or reason for choosing how many threads you run to get a true advantage. In some cases 4 does better and in other cases 2.

Top of pagePrevious messageNext messageBottom of pageLink to this message   By Burner (216.37.4.71) on Wednesday, June 06, 2001 - 09:18 am:

One last test.

I used wave filter and digital darkroom to set up a bunch of effects that kicked the render time up around 3 minutes for this simple scene.

Most of that is just post processing time. Using multiple threads actually slows the process down.

Would Altivec enchancements speed up things like motion blur and anti-aliasing?

Top of pagePrevious messageNext messageBottom of pageLink to this message   By Arnie (66.88.8.234) on Wednesday, June 06, 2001 - 11:13 am:

Motion blur and antialiasing require full rendering of extra sub-pixels, they are not post processes or 2D filters. Wavefilter is particularly slow due to the simplistic algorithms used, both in its original 32-bit incarnation and in the 128-bit rewrite.


Add a Message


This is a private posting area. A valid username and password combination is required to post messages to this discussion.
Username:  
Password: