I wanted to share some very basic performance numbers and get the team's
take on these.
First, some basic info:
- I have wmode=direct on (let's get that out of the way)
- We have a 3952 poly character model. We're using 8 different
textures/materials. 7 are 512x512, and 1 is 256x256.
- We have 2 lights in the scene.
- I've added a basic "clone" keyboard event handler that clones the
character mesh and adds the clone to the scene.
- We're trying to simulate "unique" characters in the test. Different
textures, different geometry, etc. We are trying not to take advantage of
instancing, as our application eventually will have many unique characters
in the scene.
- At the moment, all geometry in the scene is within the camera view.
Culling code is of course running, but everything is testing "in frustum".
- Hardware: Core i7 Q740 1.73GHz with 8GB RAM. NVidia Quadro FX 1800M
GPU.
The numbers:
*# Chars / # Polys / FPS Range*
1 3952 52-62
2 7904 41-47
3 11856 38-41
4 15508 32-34
5 19760 27-29
6 23712 21-27
7 27664 21-25
8 31616 18-22
9 35568 17-19
10 39520 15-18
The questions:
1. The Mesh.clone() method - does this meet the "unique characters"
requirement of my test above? Is there instancing going on underneath? It
looks like clones share geometry and textures, but I may not have dug deep
enough - the renderer itself may still create unique vertex buffers/etc for
this data when rendered.
2. These numbers seem abysmal in comparison to what I've seen in some
blog-posted molehill performance tests. I have not looked far under the
hood of these tests yet. (a)
http://iflash3d.com/performance/unity3d-vs-molehill/ (b)
http://www.nulldesign.de/2011/03/02/molehill-demo/.
3. Based on what you know about your own internal architecture, and
limitations of molehill, is there anything obvious in what i'm doing above
that would be causing the abysmal performance? Big textures?
Any additional thoughts would be most welcome.
-Dave
