@Fact Checking and I were discussing AP HW in the TSLA Market Action thread and it was not appreciated. This seems like a reasonable location, so dragging the conversation here ...
Boost / turbo speeds on pretty much all modern hardware (GPUs and CPUs) are not guaranteed and are best effort, not just due to thermal and power limits but silicon quality. Base clocks are really the only thing guaranteed (and sometimes they're not even advertised - it's not unusual for the "up to X" be the only thing mentioned on the box / PR).
Here is a reference to how different identical GPUs, as an example, can have a performance variance (the timestamp is 17m20s in case the link doesn't go directly there). That's why I was saying it's unlikely they'll assume they can use maximum rated boost speed on either the Pascal GP106 GPU or the Pascal iGPU in the Parker SoCs. They'll figure out what they can reliably get across most samples, and probably do a burn in test to ensure it (and reject the few that don't make it), so that's another few % loss on theoretical peak performance.
I expect the NN hardware will be designed for a given speed, and only that speed, no base vs boost clocks, and anything that fails at that speed is outright rejected, possibly before it's even packaged. That will give them the guaranteed performance they'll want without variance.