Most, if not all, subtests of GeekBench need less than 28 seconds, so it is quite possible for the entire benchmark to be run at an 105 W power consumption. Whenever a subtest finishes, the power consumption momentarily drops, which resets the 28 second timer.
If the computer has poor cooling, it may happen that when the CPU spends too much time and too frequently at an 105 W power consumption the junction temperature limit is reached, which triggers thermal throttling and the power consumption is reduced. This is a different mechanism, independent of the one that reduces the power consumption down to the nominal TDP after 28 seconds, or after another configured time.
Thermal throttling reduces the power consumption only enough to keep the temperature under the limit, so the power consumption may remain greater than the TDP until the 28 seconds pass.
I've done so, on many occasions, with actual power meters rather than trusting software power estimates. You really do commonly see a laptop's power consumption drop significantly ~28 seconds into a multithreaded CPU benchmark.
> (Obviously the cooling system could take it away, but part of your argument is that the cooling system is only good for 35W!).
I make no such claim that the cooling system is limited to 35W. I only claim that the default platform power management settings from Intel impose a 35W long-term power limit, unless the system builder has adjusted the defaults to account for whatever form factor and cooling choices they've made.
Perhaps you haven't realized that the turbo power limits will still kick in even if the CPU die temperature is not too hot—because they're not actually a temperature-based control mechanism?
Now I see where the disconnect was. You're right, if this was a laptop that could happen. It isn't, and it didn't.
The only relevance of the socketed nature of this part is that it is easy to put it in a normal desktop form factor where a big heatsink and possibly tweaked turbo limit settings can be used to generate misleading benchmark results. But it's not actually certain that this is what's happening; the Intel-recommended default behavior for this chip can plausibly produce the reported results—just not in any way that could be reasonably described as "35W".
Again, the thermodynamic argument is fundamental here. You're saying that a "35W" CPU is "actually" drawing power equivalent to a 125W CPU for exactly the time of a benchmark, which is several minutes. That excess would have nowhere to go! There's no reservoir to store it. (Obviously the cooling system could take it away, but part of your argument is that the cooling system is only good for 35W!).