Since the Intel e-cores still have a relatively wide decoder, e-core designs may be the part where the bill comes due for x86 in terms of decoder complexity. Sure it's only 3% of a performance core, but if you cut the performance core in half then now they're 6%. And the decoder doesn't shrink that much, Gracemont has a 3-wide decoder vs 4-wide on Golden Cove, and you still have to have the same amount of instruction cache (instruction cache hit rate depends on the amount of "hot code", and programs don't get smaller just because you run them on e-cores). A lot of the x86 "tricks" to keep the cores fed don't scale down much/any.
edit:
Intel Golden Cove: 7.04mm^2 with L2, 5.55mm^2 w/o L2
Intel Gracemont: 2.2mm^2 with L2, 1.7 mm^2 w/o L2
Apple Avalanche: 2.55mm^2. (I believe these are both w/o cache)
Apple Blizzard: 0.69mm^2 (nice)
Note that N7 to N5 has roughly 1.6x logic density scaling - so a Blizzard core would be 1.24mm^2 or roughly 73% of the transistor count of Gracemont for equivalent performance! For the p-cores the number is 82%.
This is one of the reasons I feel Apple is so far ahead. It's not about raw performance, or even efficiency, it's the fact that Apple is winning on both those metrics while using 2/3rds the transistors. It's not just "apple throwing transistors at the problem", which of course they are, but just they're starting from a much better baseline such that they can afford to throw those transistors around. The higher transistor count in total is coming from the GPU, the cores themselves Apple is actually much more efficient (perf-per-transistor) than x86 competitors.
Of course, it doesn't help that Intel lists laptop chip turbo frequencies to use either 95w or 115w and Anandtech's laptop review of one had the 12900H hitting those numbers with sustained power draw at an eye-raising 85w. That's 2-3x the power of M1 Pro and only 20-30% more performance.
That laptop also showed that cutting power from 85w to 30w roughly halved the performance. On the plus side, this means their power scaling is doing pretty well. On the negative side of things, it means their system gets worse multithreaded performance at 30w despite having 40% more cores.
A55 is probably 1/8 the performance, but something like 1/100 of the power consumption and a miniscule die area. I wouldn't want to have all A55 cores on my phone though.
Performance per die area is also relative. For example, Apple clocks their chips around 3GHz. If they redesigned them so they could ramped them up to 5GHz like Intel or AMD, they would stomp those companies, but they would also use several times more power.
What is really relevant is something like the ratio of a given core's performance per area per watt to the same value for the fastest known core.
The only interesting area for ultimate low-power in general purpose computing is some kind of A55 with a massive SIMD unit going with a larabee-style approach for a system that can both do massive compute AND not have performance plummet if you need branchy code too.