Epic's Unreal Tournament has one of the best software-based 3-D renderers that I've ever seen. For a software renderer that does as many effects as it does, it's quite fast. This is mainly due to the well-designed span buffer implementation of the renderer (which greatly eliminates pixel overdraw). It also owes its speed to the large quantities of assembly code that do the actual pixel pushing.

It's the assembly code bit that steered this project into my lap. I burned a lot of hours calculating structure offsets, poking through memory, and converting MASM assembly syntax into GAS syntax. The Linux software renderer has been folded into the main Linux source tree, and it was released into the wild with the Linux 428 patch of Unreal Tournament.