"One order of  430 BILLION  polys coming up!"

by Steve Blackmon

While doing some final memory optimizations and bug fixes to the instanced geometry handling for Brazil 1.0, I found I was able to easily render 150 million faces on a dual pentium-3 550 Xeon with 1 Gb of ram.  This sparked my curiosity so I spent a couple of evenings over the weekend working on the Large Dataset ray tracing accelerator that was already in the late design/early implementation stages for Brazil 1.1.  Very quickly I was able to exceed 1 billion faces and was quite surprised by the low render times and decided to push a little further just to see how far it would go.

Along the way some bottlenecks were discovered in scene processing, and these had to be overcome before proceeding.  Once that was done, and a few optimizations performed, I turned Irfan Celik loose (Irfan works for Blur Studio).  Exceeding 50 billion faces became trivial, and between us we tried to find a way to get even more geometry into Brazil, as Max was starting to become a problem even on a dual 1.7 Ghz Athlon.

Screen capture of the max scene used in the render of 102 Billion faces
Memory used to render 240 MB - 11min54sec - Dual AMD Atalon 1.7GHz

We finally settled on a tesselated version of the famous Stanford dragon, which was just shy of 3.5 million faces.  Showing great patience, Irfan was able to assemble a scene composed of 123,616 dragons, for a total of 430,882,852,096 faces.  The scene used 450 Mb to render at 4K, and 6.5 hours at AA 1/3 with two ray traced lights.

Just for some perspective, it takes around 3.5 hours just to load this file.  Again, this was with minimal speed optimization.  Most of the render time was spent on the back rows of dragons, where antialiasing frequently reached the maximum number of samples, and each bucket contained well over 100 million faces.  Just for fun, we let it render with skylight, as well as the two ray trace lights, which took around 12.5 hours.