Perspective-correct texture mapping requires one division per pixel. Before the advent of GPUs this was a problem because this was quite heavy to do on the CPU (especially back in the days of non-SSE single-core CPUs, when they were capable of calculating only one division at a time, requiring several dozens of clock cycles per division).
After GPUs became a thing, however, this solved the problem and games got perfectly perspective-correct texture mapping on each rendered pixel in real-time.
But this raises the question: How fast are the divisions done by the GPU when it's texture-mapping, and how many divisions is it doing in parallel? How is this done? Is there like a huge array of units in the GPU doing hundreds of divisions in parallel, like a huge SIMD unit dedicated to divisions? Or how is it done?