Not just that: you have dozens, hundreds or thousands of unit A, and of unit B, and of unit C. You test each of these units once, and then some measure of interplay.

How long would formal verification on something this big even take?

Do you mean RTL to gates formal verification?

Yes. Also, I’m assuming[a] that a GPU would take about the same as a CPU because despite GPUs having billions more transistors, they have a lot of duplicated modules compared to a CPU which has many more different modules but probably only one or two of each (per core).

[a]: My knowledge of creating hardware is very limited. So I could be completely wrong.