PDA

View Full Version : Actual geometry renders faster than instances?!



vector
06-01-2013, 12:29 PM
HI

I attach two scenes with the same group of leaves: one scene with leaves instanced and other with actual geometry.

Instanced ones took about 3 min to render a frame. The other scene (actual geometry) took about 2 minutes. Final gather Interpolated and AA. I used a I5 Pc

Am I doing something wrong? This way instancing dissapointed me :(

Vector

Sensei
06-01-2013, 05:47 PM
Actual geometry renders faster than instances?!

Did you expect reverse?
How come?

UnCommonGrafx
06-01-2013, 11:05 PM
You aren't using instances in an instance where their strength is: larg(er) data sets.

Try a scene with a million clones as opposed to a million instances; I'd be curious as to what you find.

vector
06-02-2013, 12:25 AM
I know what I get... what I didn't know you need a minimum of thousands/millions of instances to get instancing useful.

XswampyX
06-02-2013, 01:36 AM
That not what I get?

Instances -> 2m 15 secs
Geometry -> 2m 32 secs

That makes the instances faster.

http://i465.photobucket.com/albums/rr16/xXswampyXx/Instanced_Leaves_zps02c23e6e.jpg (http://s465.photobucket.com/user/xXswampyXx/media/Instanced_Leaves_zps02c23e6e.jpg.html)

http://i465.photobucket.com/albums/rr16/xXswampyXx/Actual_Leaves_zps8e3a7938.jpg (http://s465.photobucket.com/user/xXswampyXx/media/Actual_Leaves_zps8e3a7938.jpg.html)

vector
06-02-2013, 01:54 AM
Hmmmm....strange. I try at work also (not the same scene but leaves) and I got longer render times with instances, although the scene I attached gave me the times I wrote before, very different than you... not sure what happens. Thanks for taking time to test it

XswampyX
06-02-2013, 02:01 AM
No problem. I have an AMD Phenom II X6 1090T. I think the 6 cores must help with the instances?

Oedo 808
06-02-2013, 02:24 AM
On mine I get 3:36 on the instances vs 3:47 on the geometry so only marginally quicker, but still quicker as with XswampyX's result.

That was running a Q6600 at stock, whatever that is, usually @3.4GHz but stability has gone to pot since I added in some more RAM.

pinkmouse
06-02-2013, 03:08 AM
1m50s on the instances, 1m57s for the baked geometry. That's on my i7 3770 Hackintosh.

Sensei
06-02-2013, 05:27 AM
Instances rendering MUST be slower than real geometry.
Instancing requires transformation of ray origin and ray direction through matrix from world coordinate system to instance reference item coordinate system, and then going back, for every ray passing through bounding-box where is instance when there was hit.
Doing more math (instances) vs doing less math (geometry).
Turning on various caches, like GI, or new unified AA system, is fooling results. Instead of real ray hitting, there is read cache and interpolated data is returned.

vector
06-02-2013, 08:03 AM
Nice explanation, thanks... I'm not sure why I'm getting so different result. I'll try again. Thanks to all for testing it

vector
06-02-2013, 08:51 AM
Here I go again. I have to say we are using 11.0.3 in production so the test I talked above was done with that Lw version.
Although we can update to 11.5, we prefer not to do that while working on this project. But now I used it to compare and...

New test

Lw 11.5:
Instances 2'28''
Baked 2'32''

Lw 11.0.3
Instances 2'58''
Baked 2''03''

I did it twice getting similar results, without restarting the computer, alternatively

Vector

Sensei
06-02-2013, 09:16 AM
I have 149.2 sec with instances,
and 166.5 sec with geometry.
But it's ONE THOUSANDS items!
You should try baking them to ONE item..
Save all these 1000 items to one, and try again..
Just for a test, because I see they're animated.