Well, if even just one polygon was visible of a model with 130,000 polygons, the rendering speed would be very slow. The reason is that, when even just one polygon is visible, the entire model would be rendered. Try it once in the 50,000 polygon sample supplied. Turn the camera away so that it isn't visible then turn the camera just slightly so that just one polygon is visible and take note of the frame rate in each case.
This experiment was mainly to just find where the peak performance was. For more conclusive results, I'd need to test in the 1000 to 5000 polygon range and with more steps involved. The number of visible objects doesn't seem to have that big of an influence as it seems exponentially related rather than all over the place. In this experiment, there was just 3 steps to judge from with the 1000 to 5000 range (1000, 2000, and 5000). I should next go with a 10 to 20-step logarithm scale from 1000 to 10000 or something. Maybe next after this, I should find out what is the best way to optimize a model with multiple textures. That is, will 6000 polygons with 4 different skins render just as fast as the 2000 polygon model with 1 skin? A previous experiment (above in this thread) told me that it was 62% faster. Although a level designed like this would seem ridiculous for a level, there's no problems involved when it comes to nothing more than experiments. After all, it's the best way to learn! Now that I have the general peak performance area somewhere between 1000 and 10000 polygons, I'd now like to narrow it down to just this range with more precision.
Of course, these results could vary quite strongly between systems. The peak could be around 10,000 polygons (high end video card with low-end processor), or around 800 polygons (low end video card with high-end processor). 3 GHz processors seem to be about 18 months old and Radeon 9600 XT also seems to be just as old so my results could be the general average.
Why logarithms you may ask? Who'd do something linear counting by 500's to 50,000 (having 100 steps) when you're leaving out 10, 200 and other smaller values that could play a role. The actual 1/3 mark from 1000 to 10000 is 2154 polygons and the actual 2/3 mark is 4642 polygons (both of which are weird and difficult to work with), close to what I have.
Oh, and if you want to know the fps I was getting, take the polygon count for the model, multiply it by the number of visible objects, then take this result and divide it by the related column. The actual numbers given are ±2%.
