I'm not sure the difference is so great. It's all about the node about the node... (to the tune of "all about the bass" :)
I just got a 2020 MacBook Air with an Ice Lake 10th generation 10nm X64 chip in it.
It's a quad-core, newer core rev, has AVX2 and a bunch of other stuff.
It runs very noticeably cooler than a 2019 Air with a 14nm Amber Lake chip in it. Battery life is also noticeably better. When I read the spec differences I basically traded my 14nm Amber Lake Air for a 2020 Air (and also because of the better keyboard). It's a great machine.
The difference between the Amber Lake and Ice Lake core designs is not substantial and the Ice Lake has twice the cores and a larger on-board GPU, so how is it so noticeably cooler? I can fully max out all four cores of the 10nm Ice Lake and it doesn't get as hot as the 14nm Amber Lake did at lower loads! The answer is obviously the process node: 10nm vs 14nm. It's a more efficient chip at the physical circuit level.
ARM has some intrinsic power advantages over X64. The biggest thing is that ARM instructions come in only two or three sizes and it's easy to size and decode them, while X64 instructions come in sizes from one byte to 16 bytes and are a massive pain to decode. That decode cost comes in energy and transistors, but it's worth pointing out that this is a mostly fixed cost that shrinks as a percentage of the overall CPU power/transistor budget as the process node shrinks. In other words the cruftiness of X64 remains the same as you go from 22nm to 14nm to 10nm.
Other than the ugly decode path the ALU, FPU, vector, crypto, etc. silicon is not fundamentally different from what you'd find in a high-end ARM chip. A lot of the difference is clearly in the fabrication. ARM chips have been below 14nm for a while, while Intel X64 chips have lagged.
(Tangent: since the actual engine block is largely the same, I've wondered if Apple might not slap an X64 decoder in front of their silicon in place of the ARM64 decoder and make Apple X64 chips?! I am not a semiconductor engineer though, so I don't know how hard this would be and/or what IP issues would prevent this. Probably unlikely but not impossible. They certainly have the cash and leverage to muscle Intel into licensing anything they need licensed.)
If Intel gets its act together with process nodes and/or starts using other fabs who are at tighter lower power nodes, the advantage will shrink a lot. AMD already has mobile chips that are close to Apple's ARM chips in performance/watt, partly because they are fabbed at TSMC at 7nm.
BTW: I'm not claiming Apple's chips aren't impressive, and as long as they don't lock down MacOS and make it no longer a "real computer" I personally don't mind if they go to ARM64. Also: the fact that Apple's chips only get this great performance in bursts is mostly due to power and cooling constraints on fanless thin phones and tablets. In a laptop with better cooling or a desktop they could sustain that performance no problem.
I just got a 2020 MacBook Air with an Ice Lake 10th generation 10nm X64 chip in it.
It's a quad-core, newer core rev, has AVX2 and a bunch of other stuff.
It runs very noticeably cooler than a 2019 Air with a 14nm Amber Lake chip in it. Battery life is also noticeably better. When I read the spec differences I basically traded my 14nm Amber Lake Air for a 2020 Air (and also because of the better keyboard). It's a great machine.
The difference between the Amber Lake and Ice Lake core designs is not substantial and the Ice Lake has twice the cores and a larger on-board GPU, so how is it so noticeably cooler? I can fully max out all four cores of the 10nm Ice Lake and it doesn't get as hot as the 14nm Amber Lake did at lower loads! The answer is obviously the process node: 10nm vs 14nm. It's a more efficient chip at the physical circuit level.
ARM has some intrinsic power advantages over X64. The biggest thing is that ARM instructions come in only two or three sizes and it's easy to size and decode them, while X64 instructions come in sizes from one byte to 16 bytes and are a massive pain to decode. That decode cost comes in energy and transistors, but it's worth pointing out that this is a mostly fixed cost that shrinks as a percentage of the overall CPU power/transistor budget as the process node shrinks. In other words the cruftiness of X64 remains the same as you go from 22nm to 14nm to 10nm.
Other than the ugly decode path the ALU, FPU, vector, crypto, etc. silicon is not fundamentally different from what you'd find in a high-end ARM chip. A lot of the difference is clearly in the fabrication. ARM chips have been below 14nm for a while, while Intel X64 chips have lagged.
(Tangent: since the actual engine block is largely the same, I've wondered if Apple might not slap an X64 decoder in front of their silicon in place of the ARM64 decoder and make Apple X64 chips?! I am not a semiconductor engineer though, so I don't know how hard this would be and/or what IP issues would prevent this. Probably unlikely but not impossible. They certainly have the cash and leverage to muscle Intel into licensing anything they need licensed.)
If Intel gets its act together with process nodes and/or starts using other fabs who are at tighter lower power nodes, the advantage will shrink a lot. AMD already has mobile chips that are close to Apple's ARM chips in performance/watt, partly because they are fabbed at TSMC at 7nm.
BTW: I'm not claiming Apple's chips aren't impressive, and as long as they don't lock down MacOS and make it no longer a "real computer" I personally don't mind if they go to ARM64. Also: the fact that Apple's chips only get this great performance in bursts is mostly due to power and cooling constraints on fanless thin phones and tablets. In a laptop with better cooling or a desktop they could sustain that performance no problem.