An interesting observation, which is even more damning for Intel: moderately multicore CPUs from AMD are made in an effective and efficient way, while Intel, despite boasting about their advanced packaging technology, is slow to adopt heterogeneous chiplets and is paying the price of this technological rigidity. This is in addition to miserable failure with new processes.

Do you see getting the most of limited 7nm production capacity as "cheating"? The only fair comparison is in the resulting performance and TDP, and Intel remains seriously behind.

Do chiplets underperform compared to a monolithic die?

As others have said, in isolation and perfect manufacturing capability, it's absolutely true a monolithic die is better.

However, because each die is smaller, you can be more aggressive with your binning, potentially having a larger number of chips which you can bin into a higher frequency bucket.

Ultimately, it's a matter of trade-offs, and what's right in one place might be wrong in another (see, e.g., AMD manufacturing mobile APUs on monolithic dies versus all their desktop and server parts using chiplets).

We are seeing more of the industry moving to chiplets and similar solutions, notably Intel with their tiles.

All things being equal a chiplet design will underperform a monolithic die. But we've already seen the benchmarks on the performance of Milan so looking at chiplets versus monolithic is mostly about considering AMD's strategy and constraints rather than how the chips perform.

Pretty much any time you have signals going off chip, you lose out on both bandwidth and latency.

Yes.