Date: Sun, 2 Sep 2001 14:37:35 -0600 From: Mike Porter <mupi@mknet.org> To: "Ted Mittelstaedt" <tedm@toybox.placo.com>, "Sean Chittenden" <sean@chittenden.org>, "Bsd Newbie" <bsdneophyte@yahoo.com> Cc: <freebsd-questions@freebsd.org> Subject: Re: overclocking and FreeBSD stablity... Message-ID: <200109022037.f82KbZl09087@c1828785-a.saltlk1.ut.home.com> In-Reply-To: <001301c1338a$7e753b00$1401a8c0@tedm.placo.com> References: <001301c1338a$7e753b00$1401a8c0@tedm.placo.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sunday 02 September 2001 02:37 am, Ted Mittelstaedt wrote:
> >-----Original Message-----
>
> From: Mike Porter [mailto:mupi@mknet.org]
>
> >Sent: Saturday, September 01, 2001 8:03 PM
> >To: Ted Mittelstaedt; Sean Chittenden; Bsd Newbie
> >Cc: freebsd-questions@freebsd.org
> >Subject: Re: overclocking and FreeBSD stablity...
> >
> >On Friday 31 August 2001 10:22 pm, Ted Mittelstaedt wrote:
> >> >-----Original Message-----
> >>
>
> No no no that's not what I meant. It takes time for a signal to go from
> one side of the die to the other. While you have a point in that
> everything is relative to the master clock, it's very easy to have a long
> and a short trace inside of a chip where a pulse is applied to the the ends
> and at the other ends it arrives at different times.
>
> While it may be negligable because the designers compensate by holding the
> clock hi or low long enough for a hi or low bit to register, I do know that
> the chip designers are already having to deal with problems like this
> caused by the higher and higher clock speeds.
>
The amount of time a signal must be high or low is tied to the clock; the
signal must be applied until the clock state transitions (for DDR systems) or
until the clock state transitions low->high (or high->low...i forget which,
but in any case it is only one of them, DDR systems switch at BOTH
transitions, effectively doubling the clock rate) for "normal" (ie, intel)
systems. Granted that there is a certain finite amount of time for silicon
to switch states, which is why, for example, your 10ns SDRAM won't work in a
system above 100Mhz (which is why PC133 RAM is 8ns parts). (This is 1/2 of
the reason your old pc100 or pc133 RAM won't work in a DDR system, too) This
finite time varies a great deal on things like process size (which is one
reason smaller process size can yield higher clock speed, even without the
physically smaller form factor), and quality of the silicon involved. This
is why it is at least theoretically possible that Intel tests chips as
alleged; a part that due to silicon impurities is not capable of switching a
500Mhz, may be capable of switching at 400Mhz. If it is tested and passes at
400Mhz, it doesn't matter that it failed at 500Mhz...even in todays litigious
society, I would be surprised if someone was able to win a case like that, SO
LONG AS they have records that it passed at 400Mhz.
> This applies to die sizes much smaller than those
>
> >currently in use and to clock speeds much higher than those
> >currently in use.
>
> Um, not exactly because a square wave (which is what a clock is) generates
> lots of nasty harmonics at the transition (they call it the leading and the
> trailing edge) some of which are much higher frequency than the frequency
> of the square wave itself. That has to be taken into account as well in
> the design of the chip.
>
True, but that affects the clock speed and synchronization of signals within
the chip, not the amout of time it takes fro a signal to cross from signal
origination to signal destination, which is a factor of clock speed (that is,
the wavelength of the frequency) and the distance. And there is no easy way
to figure what the MAXIMUM signal length is without knowing a great detail
about the chip design. The simple, obvious ploy of "one side ofthe chip to
the other" doesn't apply becuase I can't think of a single signal on a
Pentium class chip that would have to traverse the entire width of the
silicon; they design things so that the signals travel very short
distances...I think the longest I have ever seen stated was something like
1/4th the distance across the chip; THAT is the minum threshold. Once we
reach a point where the length of a clock cycle is smaller than the length of
the longest signal path (for argument, lets say 1/2 of the die size) that is
when you reach the theoretical limit of today's technology. But with
shrinking die size, that barrier keeps increasing. At that point they will
either have to come up with a new design (optical chips, as you mention, are
very promising) or figure out a way to buffer things inside the chip (not
terribly difficult in principle, the instruction (or instruction/data
pipelines on a sparc) are exactly that, but syncronizing the various parts of
the chip to each other would be a nightmare *I* wouldn't want to have to
tackle.
> But before that time we will run into a different issue - that is that
> you may be able to get the signal propagated properly, but you must hold
> the signal high or low a certain amount of time to get the semiconductor
> receiving it to actually register a logic high or low. The faster the
> clock the less time you have to hold the signal at a given state.
>
That is a different issue. I don't know the details, but as I mentioned, so
far, they aren't to this threshold yet, or DDR chips wouldn't work. And
isn't the P4 supposedly able to do QDR? Again, it would appear that the
propagation WITHIN the chip is much less of an issue than the propagation
into and out of the chip (which have much longer signal paths...again, once
you exceed the distance that can be covered in a single clock, you HAVE to
buffer somewhere.
<snip possible urban legend story>
> I've heard this story before, usually from people wanting to overclock, and
> I really question that this actually happens.
>
I actually got it from Tom's Hardware guide. It has been there for a *long*
time, and I would imagine that Intel has had ample opportunity to get them to
take it down *if it wasn't true* ("umm...if you don't take it down, we won't
ship you any more pre-release chips" ). This story has been circulating a
long time, and Intel has surely had many many opportunities to deny it. The
cynical side of me is inclined to argue that they might not *want* to deny
it, as it leads to more people overclocking, thus more people frying chips,
thus intel sells more chips. However, Intel has almost certainly lost
revenue from overclocking; look at the pains they took with the PentiumII to
try to "lock" the chipset clock and multiplier to settings inside the chip.
(and the p3, too....I don't know about the p4 but I suspect the same to be
true there). AMD also, with the athlon's, although that is fairly easy to
defeat on the Thunderbird/Duron chips.
> There's no doubt that Intel tests the CPU's before packaging them, as the
> packaging costs money and it's stupid to package a failed chip.
>
> But, beyond that, I'm not so sure. Every time I've asked anyone who
> actually WORKS in process control at Intel (and keep in mind that one of
> the plants that Intel makes Pentiums at is located about 3 miles from where
> I work) I've been told that this story is an urban legend and a pile of
> baloney. The problem is that if Intel sets up a production line to make,
> for example, 500Mhz CPU's, then if the chips don't pass the test then
> there's too much liability to retest at a lower speed, hoping to get a chip
> to pass. Instead they throw away the CPU. Consider that if they DID pass
> the CPU at a lower speed then they would in effect be selling a CPU that's
> a failed part. Remember that Intel's chips go into a lot of other gear
> besides consumer PC's that have no warranty, there's military gear and
> medical gear a bunch of other stuff where they are held to a lot higher
> warranty standard.
>
Why would you build a run of a product that is capable
of 800Mhz ( I think that is what the current Celerons can run to), then
retool everything to build a 750Mhz part that is *identical in all respects
except rated clock speed* to the 800Mhz part? I don't know what test Intel
uses, but when I worked for 3Com, we had a big machine that would take
completed PCBs (each PCB had four cards on it, be they network cards, modems,
whatever) and before cutting the assembly apart, would test each piece.
Any failed test, would rerun using a set of parameters that *could* be
specified to be different from the original test (we never used it that way,
but the test setup included five configuration files, one for each possible
failure...I believe, although I never got that far into it, that if we had
created a sixth file, it would have run the test six times instead of five,
but they figured if it failed four tests, it probably was a failure. The
point is, if Intel uses a similar tester (this was an HP system, which I
believe to be more-or-less an industry standard for in-circuit testing of
parts), then all they have to do is change the clock parameters for the
second test, the third test, etc.
Do they actually do this? I can't say, since I don't work for Intel. I
suspect that those that do can't say either, unless Intel decides to
officially comment one way or the other. I am pretty sure that they don't do
this with the top-of-the-line chips, as they want to get the maximum number
of higher-clocked chips out there, to support the R&D cost. It is also
likely that AMD at leased *used* to not test the way, for much the same
reason (as an uderdog, they want to get the maximum number of
highest-performing chips out there).
> Consider that if someone dies on the operating table as a result of a
> failed Intel CPU then if their lawyers find out that the CPU had failed the
> test that it was _supposed_ to be produced for, there would be hell to pay.
>
As mentioned before, I don't necessarily agree with you here becuase there
isn't a *supposed to* in the production value. You build the chips, and
whatever they test to is what they are *supposed* to be. Due to variations
in the quality of the silicon, even from wafer to wafer sliced off the same
silicon log there is no way to know for certain what quality you will get.
given the phenomenal cost of equipment, silicon, and R&D, I seriously doubt
that even Intel has any great desire to waste more than they absolutely have
to. Besides, how could you ever PROVE that the chip was *supposed* to be
800Mhz? the 800Mhz chip is (or is supposed to be) in all respects the same
as the 700Mhz chip that I bought, except rated clock speed. And guess what?
I have been running at 800+ MHz for almost a year, and I rarely shut my
computer down. My hard drives get hotter than my CPU.
> I think that the reason that overclocking works is that it's standard
> practice to heavily derate electronic parts. This is really a cost issue.
> For example, it costs a fraction of a cent more to manufacture a diode that
> passes 1 amps before burning up rather than one that burns up at 1/2 amp,
> so as a manufacturer your really stupid to not make 1 amp diodes and mark
> them as 1/2 amp, 2 amp diodes and mark them as 1 amp, and so forth.
>
While it sounds good in theory, I don't see it happening in practice. I have
by 700Mhz Celeron overclocked to 800, no problems. I didn't even have to
increase the core voltage to make it work. At the same time, I have an older
200Mhz Celeron that if I got over about 225Mhz refused to even boot. Same
hardware, motherboard, disks, etc. But one chip I get a 114% performance
over rated, with no problems I have found, the other chip I can get 112%, and
not only does it fail, it fails miserably, won't boot or anything. Clearly,
it isn't *just* a deration issue, or you would expect similar results from
similar hardware. For what its worth, I have seen similar results from AMD,
IBM, and Cyrix...some chips overclock fine, to significant amounts, others
refuse to overclock more than a couple of percentage points (about enough to
accomodate the performance cheating of some motherboards which CLAIM 100Mhz
clock and actually deliver 105Mhz or so...oddly enough...Tom's hardware did a
test, and one of the overclocked mobos actually performed more poorly than
its more conventional counterparts. go figure...overclocking doesn't ALWAYS
work!) (and even then, some of the problems people have with certain
motherboards may be do to exactly that...swapping out a CPU might fix the
problem...)
> In some instances, deration is built into the law itself - try getting a
> licensed electrician to install a 15 amp circuit that's intended to feed
> a device that draws EXACTLY 15 amps, for example.
>
I have to be difficult, but I have never heard of such a device that draws
exactly 15 amps all the time, never more, never less. On the other hand, its
hard to find much of anything for 15amps....ts pretty much all 10 or 20,
except in cars.
> With chips, they probably design the production line to make 600Mhz CPU's
> then start the line and mark every single part coming off the line as a
> 500Mhz CPU. Consider that these CPU's are going into really grungy
> motherboards, have you ever measured the speed of a typical motherboard
> with a test instrument? Even without deliberate overclocking it's not
> uncommon for garbage-grade motherboards to vary 5% or more on both CPU
> clock and voltage. Doing it this way protects Intel (or whomever) because
> in a given run you may have parts that will work up to 700Mhz and some
> parts that won't work better than 525Mhz.
>
Again, there is no real difference between a chip running at 600 Mhz and a
chip running at 500Mhz, so why go to the effort of setting up the line to
produce 600Mhz parts? Isn't it easier to simply test the parts as they come
out of production, and label them according to what they can do? And isn't
it more effective? If you can sell a part rated to 700Mhz, wouldn't you want
to capture that revenue? Granted tends to argue for testing each chip
individually, but there are economies of scale involved, and ease of access
to the parts to test....
> >The other wrinkle in this scheme is that Intel is completely free, if the
> >demand is there, to remark their OWN chips to a LOWER speed. So if demand
> >spikes for a 300Mhz Celeron, and they have a pile of 450Mhz Celerons
> > sitting on the shelf, there is nothing illegal, immoral, or fattening
> > about calling them 300Mhz Celerons and pricing them accordingly. (after
> > all, they passed as
> >450Mhz celerons..and don't forget, any given chip in the lot of 450's has
> > a one-in-ten or so chance of being capable of much faster speeds than
> > even 450Mhz).
>
> Once more I've been told this is a crock, because Intel cannot afford to
> have a
> "pile" of CPU's sitting around, the economics of chip prices are such that
> every day that a completed CPU sits there and isn't bought, the company
> loses money on it.
>
Exactly my point. If they begin to accumulate a "pile", they have a very
strong incentive to mark them down until they sell. If they accidentally
build too many 800Mhz celerons, are they going to wait until they sell, or
are they going to sell them as 700Mhz celerons? And don't frget that if they
made too many 800Mhz celerons, they probably also didn't build enough of the
700Mhz ones! (they don't, after all, have infinite capacity to build chips,
if they are building an 800Mhz chip, they AREN'T building 700Mhz chips)
> Now, that's not to say that they don't do this with _old_ CPU designs, I
> understand for example that there's still 8088 family chips being used in
> embedded systems. Undoubtedly the tolerances on an 8088 being manufactured
> today are lightyears better than they were 15 years ago, so you would
> probably be able to overclock one.
>
embeded systems and SCSI drives. Though usually a SCSI drive uses an 8086
becuase they have higher data throughput than the '88's. There are also a
lot of the '186 and '188 in use there, those never caught on very well in the
consumer market. They are also dirt cheap.
> It would really be interesting if someone who actually WORKS in process at
> Intel would be willing to officially confirm or deny what the real story
> is.
>
I second the motion! <(}:
> One of the problems that I think we have with the computer companies like
> Intel and AMD is that they know damn well that if they would just change
> the computer BOX design, we could get a whole lot of performance a hell of
> a lot cheaper. But instead they seem locked into this idea that no matter
> how fast the CPU runs, it absolutely MUST do it at whatever temperature you
> can obtain with a big blob of aluminum stuck to it and a fan, with
> room-temperature air blowing across it. There seems to be zero interest in
> investigating changing the operating temperature of the computer itself -
> at least for consumer designs that is.
>
There is a great deal of interest in this, just not from Intel or AMD. Intel
and AMD make money by selling people faster chips, not by selling them ways
to make existing (slower) chips faster. There are a couple of interesting
active cooling designs that are quite efficient. The most interesting rely
on thermoelectric properties of some metal compounds (apply an electric
current to one piece of metal, and the other becomes colder, often
significantly so (this is the inverse of the effect that witht he same to
metals, if one is hotter than other, will generate electricity. And that's
as far as my knoweldge of that extends). The least promising includes
hooking up a garden hose to your CPU tower. The idea of all that water
inside my case doesn't make me very happy, though it supposedly works quite
well.
mike
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200109022037.f82KbZl09087>