I dont think there is strong connection between a larger instruction set, and performance.

Using a bad example, let's say x86-64 and A11 can both add numbers up to 3 digits. The a11 can do that faster than x86-64. But I want to add a 5 digit number. The x86-64 can do that with one instruction, but the A11 has to break that 5 digit number into multiple 3 digit numbers in order to process it.

From my understanding, which might be wrong, Geekbench only tests what both can do. Comparing a pickup truck and a sports car you need to look at more than just how fast they can do. You can both move furniture, but the sports car will need to do a lot more to be able to what the pickup can.

On the other hand, Apple hasn't been shy about putting fixed function hardware on the Ax chips to eke out the maximum performance for a given power budget. They were much faster at adding h265 decode/encode than Intel, for instance. I would imagine they'd spend their (larger) transistor budget on a lot of this type of dark silicon that would assist in speeding up common operations.

Geekbench doesn't benchmark single instructions.

This doesn't make any sense.

Yes, Geekbench is a terrible CPU benchmark, but both of them are Turing complete, so there is no functionality difference. And complex instructions are broken into uOps in x86 too.

You are 100% correct.