The quality of the CMOS manufacturing process is the main factor that determines the multi-threaded performance of a CPU when it is constrained by a power limit.
The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design.
When AMD has transitioned to TSMC in 2019, they have instantly got a much better MT performance per watt than Intel, so the TSMC "7 nm" was already better than anything Intel had in 2019.
At that time, "7 nm" was already an old process for TSMC, having been used by Apple since 2018. Therefore, at the latest in 2018 TSMC was well in advance over Intel.
So what you said, that "TSMC only caught up to Intel's process in 2016. The playing field remained pretty even over the ensuing 4 years." is only partially correct. No later than 2018, i.e. no later than 2 years after 2016, "the playing field" was no longer even.
Perhaps TSMC was already above Intel long before 2018, but there are no older products in direct competition that would allow us to judge whose CMOS process was better.
> The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design.
I think this is an overstatement in a couple of ways:
> unlikely scenario
We're still less than a decade removed from the final descendants of Bulldozer. We can't directly compare it to Zen because of the process differences, but Bulldozer sucked in several ways that weren't related to its 28nm process.
> made a far from optimal design
AMD and Intel continue to release improved x86 architectures. Alder Lake introduced heterogeneous cores to x86. Zen 5's dual decode path is less visible, but an innovation in its own right. If there is a long-term local max in x86 design, we haven't found it yet.
That's not to say flagrantly bad design is impossible (again, Bulldozer) - it's only to say that there are indeed two avenues to making a CPU better. Even if one avenue (process) sets the hypothetical ceiling for the other (architecture), we haven't found that ceiling very often for any particular process node. Incremental improvements are a routine part of any uarch, even when the process remains the same.
The design of the CPU matters for this quality criterion only in the unlikely scenario when the design team was not competent and they have made a far from optimal design.
When AMD has transitioned to TSMC in 2019, they have instantly got a much better MT performance per watt than Intel, so the TSMC "7 nm" was already better than anything Intel had in 2019.
At that time, "7 nm" was already an old process for TSMC, having been used by Apple since 2018. Therefore, at the latest in 2018 TSMC was well in advance over Intel.
So what you said, that "TSMC only caught up to Intel's process in 2016. The playing field remained pretty even over the ensuing 4 years." is only partially correct. No later than 2018, i.e. no later than 2 years after 2016, "the playing field" was no longer even.
Perhaps TSMC was already above Intel long before 2018, but there are no older products in direct competition that would allow us to judge whose CMOS process was better.