KeyJ

@rygorous Yes, that's what I meant: The individual parts aren't necessarily rocket science, but it's a lot of busywork to do just to sample a frickin' texel (or, more likely, a bunch of texels, in SIMD fashion), with several special-case scenarios that need to be handled as well.

@rygorous Oh, absolutely. Doing a 16x anisotropically-filtered lookup in an ASTC-6x5-compressed cubemap texture with sRGB-to-linear conversion is nothing anyone's ever wanting to do in software. (Except probably as a "fun" exercise, or as some reference implementation. But not in realtime.)

@rygorous @mogwai_poet Good point. I couldn't find data on whether the TNT2 did single-cycle trilinear. So, yes, if you consider all the bells and whistles the RE did "for free" (using a *very* technical definition of "free", considering its price tag ;), like trilinear and MSAA, you'd probably come closer to a GeForce 2 indeed.

@rygorous @mogwai_poet I'm not so sure about that. The paper mentions a peak of 240 Mpx/s, which is almost exactly what a TNT2 could do. A GeForce2 is roughly 4x as fast, in terms of raw throughput.
The only thing that sets the RealityEngine apart is the absolutely unhinged memory size and bandwidth, which could afford them things like always-on MSAA, and the extremely overpowered geometry processing stage.

@rygorous "Every FG had a copy of the entire texture memory" was also true for the 3Dfx Voodoo 2 (or the rare dual-TMU Voodoo 1's, for that matter). The entire Voodoo Graphics architecture looks like a massively downscaled RealityEngine to be honest, which is of course no coincidence, considering who designed it.