In terms of boot time, I can imagine there are situations where it might be important

Sure, like in the product I mentioned before. It combines the ARM Linux system for the user interface and an ARM-based microcontroller for the product's main functions. The microcontroller has run through its bootloader and the power on self test and is ready to work before the kernel passes control to /sbin/init (no RAM-disk involved, the system boots directly from flash).

I would have liked to strip some unneeded features out of the kernel (why should it search and wait for unused or non-existing hardware), but alas, Yocto won't let me do this with a sane amount of work. Well, I could build my own kernel out of the Yocto tree and replace the vendor-provided one. But again, this takes some time. Likewise, my initial idea was to get rid of all init scripts, instead use some tiny init and daemontools to start all required software - essentially, X11, the main application, and during development, network, ssh, and a serial console. It could be done, and I'm sure it would reduce the boot time by a few seconds. The existing boot time is good enough, not as fast as it could be, and nobody complained. So it is unlikely that we'll invest time (and money) to reduce the boot time.

While the main product is powered, but the user interface is not needed, we can switch the entire display module to a deep-sleep mode. It keeps its memory content, but stops the CPU and cuts power to the LCD and most other peripherals, saving significant amounts of battery power. Waking up the display module takes about one second. When running on mains power, we don't even use the deep-sleep mode, but just switch off the LCD. Switching it back on is almost instantaneous. Sleep and display off modes are the common case, not booting. So the entire product appears to start within about one second or less most of the times, and in perhaps five or ten seconds when it was really switched off.

Not perfect, but good enough.

Alexander