It sounds reasonable to me. Around 1994/5 I was doing perspective-correct texture mapping on a Sun SPARC (with real-time dithering too), an Atari Jaguar and a

It sounds reasonable to me. Around 1994/5 I was doing perspective-correct texture mapping on a Sun SPARC (with real-time dithering too), an Atari Jaguar and a 486 DX4/100. 16k/sec polys on the DX4 (which is only ~500 at 30fps) sounds modest to me. (We had technically more texture-mapped polys than that, but they had geometric constraints with each other so they weren't all independent polys and shouldn't be counted the same way.) I was surprised, coming from earlier processors including the 386, to find floating-point was sometimes faster for calculations for things like texture UV interpolation, instead of integers (fixed-point), x86 sub-register tricks and hyperbolic-Bresenham or line-at-infinity techniques to avoid division for perspective.