Comment by rasz

Because GPU developers started to be less lazy :)

No T&L meant everything was culled, clipped, transformed and per-vertex divided (perspective, lighting) on CPU.

Then you have brute force approach. Voodoo 1/2/3 doesnt employ any obvious speedup tricks in its pipeline. Every single triangle pushed into it is going to get textured (bilinear filtering, per pixel divide), shaded (lighting, blending, FOG applied) and then in the last step the card finally checks Z-buffer to decide between writing all this computed data to buffer or simply throwing it away.

It took a while before GPU devs started implementing low-hanging fruit optimizations https://therealmjp.github.io/posts/to-earlyz-or-not-to-early...

Hierarchical-Z, Fast Z clearing, Compressed Z buffer, Compressed Textures, Tiled shading. It all got added slowly one step at a time in early 2000.