[Germont]

The I/O power consumption goes down very significantly when all the die-to-die interconnects have a small and well-defined capacitive loads.

And I suppose that having all heat sources right next/on top of each other forces you to spend design effort to lower power consumption ;)

[variaga]

It's true that die-to-die IO is lower power than package-to-package IO because of the lower capacitive load (which comes mostly because you don't need ESD protection inside the pkg) and shorter trace length, but it's still much, much higher power than the "No IO at all" case of a monolithic IC.

I've worked on the "MCM or monolithic" decision making process and the only way MCM ends up lower power is if it lets you optimize the silicon-process selection of the different chiplets and the savings from _that_ outweighs the power cost of the die-to-die IO.

OP is right - "reduced power consumption" is not an automatic benefit of chiplets. it only happens in select circumstances, which are (a) low enough bandwidth interface between the chiplets that the extra IO power cost isn't too onerous and (b) chiplets allow you to move enough of the logic to a better process node than you could use for a monolithic chip (why not move the whole monolithic chip to the better node? because you've got some critical function that is 'stuck' at an old node) to more than offset the power cost.

[bee_rider]

Out of curiosity (all I had is a VLSI class so I’m always interested to hear from folks who actually ended up working in that field), do chiplets have better power gating than monolithic?

It was a throw-away at the end there. I’d assume a whole chiplet could have its power cut completely. But I don’t know if anybody actually does that.

[variaga]

>do chiplets have better power gating than monolithic?

like a lot of stuff, the full answer is "it depends", but mostly - no. Chiplets and monolithic ICs get about the same power gating quality/efficiency (whatever your metric is).

>I’d assume a whole chiplet could have its power cut completely.

You _could_ do this - put a chiplet on an isolated power plane in the package and then cut the power to the entire chiplet. But that is IMO rare because (a) you've just moved the problem somewhere else - the power gates now need to appear on your board, not on the PCB - and (b) frequently you don't want to completely cut the power to a whole chiplet..

For instance you may want to keep some data resident in internal memory as the (power/time) cost of re-writing that data after a "full" power down may exceed the extra power saved between the full/partial power down. You'll end up doing a pretty tedious analysis of how long does it take to initialize the chiplet and how long is it likely to be powered down vs. the different power consumption in the different modes.

Anyway, on chip power gating is very good - you really can turn the power off 100% in the gated region - at the cost of extra silicon area for an isolation band around the gated region and level shift/isolation cells for all the I/O into/out-of the gated region - all of which you also have to do in the "power down the whole chiplet" approach.