I think the main point is to use 32bit on new nodes too. Using 32bit instead of 64bit is always more energy efficient unless perhaps you really need 64bit for your application, which most applications don't.

What I think will happen is that the huge amount of money that goes into developing mobile phone SoCs, which are clearly trending towards 64Bit while still requiring low power consumption, will cause that technology to trickle down and be used in ever lower cost markets as time passes and those initial costs have been recouped. The Apple M1, as a system, is clearly crazy power efficient, even at 5nm. And again, if you need an ultra-low power sleep mode, because you only wake up once a day to send some sensor data out, its probably not running Linux in the first place.

You can't make chips arbitrarily small, because you end up wasting more of your wafer for the area between the dies and for the wire bond pads. If the chip size stays the same but the CPU gets smaller, most of the chip is for off-core components and the CPU has less impact on the total cost and power consumption.

You could use more complex CPU cores, or simply more cores, for better performance, but then you also need a faster memory interface (wider buses, LD-DDR4+ instead of DDR3, ...) to make actually use of the performance. These lead to higher memory capacity as well, but then you can't actually use the available memory as you run into 32-bit addressing limits.

I wouldn't rule out shrinks of existing 32-bit SoC families to 22nm or below to lower cost once those processes get cheap enough and there is still demand for compatibility, but there is a good chance that 28nm SoCs is where 32-bit ends. (note that I'm not talking about the dozen or so additional ARC/Xtensa/RV32/Cortex-M/... microcontroller cores on high-end SoCs, as those are not the ones running Linux).

I could not find any actual paper with real use case comparison. Have you got one?

Sure, if you use any kind of 32bit mode then you'd use less energy in theory. But if we leave the responsibility to the software developer, then we all know what happens :)

The sweet spot for power efficiency has apparently moved from 28nm to 22nm, which uses less energy than either 28nm HKMG or 14nm FinFET and is also closing in on 28nm on cost.

There no mainstream 32-bit cores on 22nm or below (yet), so there is a good chance that the coming generation of low-cost 64-bit SoCs on 22nm will beat all 32-bit chips on performance, power consumption and cost.

that's not entirely true. they consume less energy in run mode, but in deep sleep the leakage currents are higher than with larger nodes. this is relevant for ultra low power applications, but those don't run linux anyway.