From: Linus Torvalds <torvalds@linux-foundation.org>
Newsgroups: fa.linux.kernel
Subject: Re: [patch] CFS scheduler, -v8
Date: Fri, 11 May 2007 16:52:21 UTC
Message-ID: <fa.oZhj8hj7kSDLnitsqrEJcRJN+RE@ifi.uio.no>

On Thu, 10 May 2007, Pavel Machek wrote:
>
> Also notice that current cpus were not designed to work 300 years.
> When we have hw designed for 50 years+, we can start to worry.

Indeed. CPU manufacturers don't seem to talk about it very much, and
searching for it with google on intel.com comes up with

	The failure rate and Mean Time Between Failure (MTBF) data is not
	currently available on our website. You may contact Intel®
	Customer Support for this information.

which seems to be just a fancy way of saying "we don't actually want to
talk about it". Probably not because it's actually all that bad, but
simply because people don't think about it, and there's no reason a CPU
manufacturer would *want* people to think about it.

But if you wondered why server CPU's usually run at a lower frequency,
it's because of MTBF issues. I think a desktop CPU is usually specced to
run for 5 years (and that's expecting that it's turned off or at least
idle much of the time), while a server CPU is expected to last longer and
be active a much bigger percentage of time.

("Active" == "heat" == "more damage due to atom migration etc". Which is
part of why you're not supposed to overclock stuff: it may well work well
for you, but for all you know it will cut your expected CPU life by 90%).

Of course, all the other components have a MTBF too (and I'd generally
expect a power supply component to fail long before the CPU does). And
yes, some machines will last for decades. But I suspect we've all heard of
machines that just don't work reliably any more.

			Linus