Assume we have a perfect cooling solution that can always keep the circuit at room temperature, you'll still need to increase voltage to get higher frequency. Since the processors run at such high frequency, the time it takes a transistor to switch state becomes nontrivial, if the switching time is too long, at the end of the clock cycle, the sequential circuits that make up our processor will have wrong result read from their combinational logic paths (propagation delay), which will result in instability like glitches or shutting down. Unless you can cool it down to some super low, sub-zero temperature (where the conductivity becomes much higher and even superconductive), this remains true.
By "normal" I was referring to most RDNA GPU we have seen so far, most of them operate at sub 2.0 GHz clock frequency. If you look at
this Reddit user's analysis, you can see to increase the sustained clock from 1.75 GHz to 2.07 GHz, the voltage increases by about 30%. I guess "normal" voltage thereby refers to
V <= 1000 mV
.
Welp I guess I should've phrased it better. I think they will test each chip individually to assign it a best base voltage.