Not at all, real-time system guarantees latency for certain operations. Meaning that RT process will never take longer than the specified amount of time to run because if it does the system has failed, it's all or nothing. Scheduling latencies in normal use OSes are much less of a problem, a user will never notice if a process takes an extra 1-100ms (or sometimes higher depending on the operation) to do it's job.
Mostly. The only time latency matters in a desktop operating system right now is when your doing live audio recording and realtime audio proccessing. I've read that the minimal possible amount a delay person can percieve is around a 10msec delay.
As a digital audio workstation it is important that a operator be able to match beats and line up different tracks and different audio sources properly for a recording. You may have somebody on a midi controller plugging away at a keyboard which is then making noises using software syths on your system, then anothe person may be playing a guitar, another on drums and so on and so forth. Maybe even this is done live, who knows. For recordings then you don't want to have everything out of sync, or the buffer on your audio I/O running out and you get pops and squells and such.
By default Windows audio system uses what is called 'kmix' which is a software mixing subsystem for win32, I beleive. (which is also why it's rare to find cards that still do hardware mixing) By having this extra layer between audio cards and the software then about the best you can expect is upwards to 250 or msecs, where you can drive the system and not get audio artifacts. Or soemthing like that. That is why you have ASIO audio drivers that bypass the normal Windows audio subsystems and allow audio applications to access the audio I/O on the comptuer directly. Still though you will see people telling users to disable things like virus scanners and such in XP to avoid heavy disk access which will cause buffer underuns in the audio hardware.
Of course Linux didn't have software mixing by default until recently. (of course nowadays dmix mixer is enabled by default, but is easily bypassed by applications if they feel like it) ALSA drivers allowed direct access, but still there has to be support in the hardware for effective low latency audio work, so nice cards are still nice. With current 2.6.18 kernels and such there have been enough realtime-like features added that you can get pretty reliable operation of maybe 100msec or so, which is good enough for most things. Older 2.6 kernels, even with preemptive features, I don't think realy do a very good job. It depends on the actual kernel version though. With audio-specific Linux distributions you have things like Ingo Molnar's realtime-preempt patches used in the kernel by default which allows reliable operation of under 10msecs on fast machines. (and of course Ingo's patches are highly unstable and bits and peices are being incorporated into 2.6 vanila kernel as time goes on)
Now this is were OS X is pretty good when compared to Linux and Windows. They have their 'Core Audio' system which seems like is specificly designed for using OS X and a digital audio workstation, which is obviously one of OS X's common uses. Low latency operation is pretty reliable (I am guessing under 30msecs reliably, but I have no personal experiance. People brag that sub 10msec is possible, but I doubt that is possible on a heavy cpu load.) and is actually something nice about OS X that makes it desirable for this sort of thing.
Thats not to say that XP or Linux can't be usefull when compared to OS X, it's just that OS X does it by default which would require special drivers in XP and special kernel in Linux.