Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Introducing "No-Moore's Law"

Author: Robert Hyatt

Date: 08:45:18 02/28/03

Go up one level in this thread


On February 28, 2003 at 01:36:00, Jeremiah Penery wrote:

>One interesting place to ask these questions might be at www.realworldtech.com
>(specifically, post a new message at:
>http://www.realworldtech.com/forums/index.cfm?action=forum&roomID=13 ).  There
>are people who read/post there from many relevant companies (Intel, AMD, IBM,
>etc.) who have huge knowledge of this stuff.  I'm sure you can get excellent
>answers there.
>
>On February 27, 2003 at 22:45:31, Robert Hyatt wrote:
>
>>If you have time, I want to ask a very precise set of questions, since we keep
>>going around and around with the non-engineering types here...
>
>Somehow, I get the impression that you're just skimming what I write, assuming I
>must be saying what you think I'm saying.

I'm not skimming anything.  You said "the chips will run much faster but they
underclock
them when they are released..."

I don't believe that, from specific questions to two long-term engineering
friends of mine,
one of whom taught in a CIS program with me for 12+ years before going to work
in the
processor design group at Intel.  Another is a former student (early 70's) that
went to work
at TI doing the same thing.


>
>>Here they are:
>>
>>(1) When you design a chip for a particular fab process, do you have a pretty
>>accurate idea of how fast it is going to clock?  To clarify, some think that you
>>crank out the design, run it down the fab line, and then if it will run at (say)
>>5 ghz, you will actually clock them at 3ghz to avoid pushing the envelope too
>>far too fast.  The engineers I have talked with dispute this with a laugh, but
>>I thought I'd give you a chance at this as well.
>
>When Intel shrunk the P4 from .18um to .13um, the processor speed first released
>in .13um was the same as the top-end part from .18um - 2GHz.  It's laughable to
>think that process shrink wouldn't give them quite a bit of frequency headroom
>from the very first, even given the immaturity of the process.

It isn't so "laughable" to me.  First, it might take time to get the fab process
tuned up to reliably produce faster parts.  But reducing the die size has other
advantages, including lower voltage and lower heat, at the same clock frequency,
so there is a reason for going there..



> Even on a very
>mature process, near the end of a core's lifespan, they _have_ to leave some
>headroom at the top, or you get processors like the 1.13GHz P3, which was pushed
>right to the limit, and suffered for it.

I don't disagree.  I simply claim the "headroom" is not very large, and it is
intentionally
made just larger than the expected variance in the parts off the fab line, so
that _all_ will
run at that speed reliably.  10%  Maybe.  25-50%?  not a chance in hell...




>
>>(2) when you design a chip for a particular process, do you have a good idea of
>>how fast it will run, or do you "wing it" and run a few off and test them to
>>see what they can clock at?  Again the engineers I talk with say that they
>>know in advance what it should run at and they run it there.
>
>There are a ton of variables that can affect the clock speed attainable by a
>certain chip design on a certain process.  If you really feel like discussing
>specifics, I can do so.  Here, I will only say that there are things beyond the
>path timings that can affect attainable clockspeed.  Thermal issues are a big
>deal for this.  Even the packaging used can affect clock scaling of a particular
>chip.

I don't believe I have said otherwise.  Also the very process of depositing
material to
build pathways is not precise, leading to variable widths and resistance/etc.
That's a
part of the fab process.

>
>They can calculate all they want, but they still have to test to make sure
>something beyond the scope of their calculations doesn't change the results.

Certainly, but I would maintain that _most_ of the time their calculations are
dead
right.  With an occasional glitch since no process is perfect.



>
>>(3)  Is there any science in the process of designing a chip, or is it a bunch
>>of "trial and error" operations?  IE in Intel's "roadmap" they are discussing
>>plans for the next 18 months or so, with predicted clock frequencies.  Are they
>>able to simply take today's chips and "crank 'em up" after a year, or are they
>>changing the fab, the dies, etc to make the chip faster.
>
>Of course they tweak the manufacturing process over its lifetime.  Processor
>cores are tweaked less often.  I ask again, do you seriously think that when
>Intel went from .18um Willamette P4s at 2GHz to .13um Northwood P4s that they
>couldn't increase the clockspeed, even given the immaturity of the process?
>


That is exactly what I believe, yes.  There is a _big_ gain if you are 2x faster
than
your competitor.  Just like everyone here looks at the SSDF list, and buys the
program
at the top, even if it is just a tiny bit higher in rating.  What would they do
if the top
program were 200 points higher?  Buy fewer?  I don't think so, it would create a
greater
demand for _that_ program.

When I purchase machines, I look at the clock speed, and I benchmark.  If one is
clearly
faster, that's what I choose.  If it is _much_ faster, I'm likely to buy
several.



>>I hope you see where this is going.  I believe, from the engineers that I
>>know (and that is not a huge number but I know enough) that this is a very
>>precise deal.  I designed stuff many years ago using TTL and CMOS stuff, and
>>the "books" were my bible for doing this, telling me _exactly_ what the gate
>>delays were for each type of chip (ie LS, etc.)
>
>Modern MPU manufacturing is FAR removed from designing small-scale TTL/CMOS
>stuff.  There are way more factors involved in clock scaling potential than just
>the gate delays, which themselves are determined by several other factors (e.g.,
>thickness of the gate oxide layers - thinner layers allow greater clock
>scaling).
>

Actually it isn't.  Silicon compilers do a lot of the work I had to do by hand.
Summing
gate delays.  Doing the routing.  And in my day, an error was a terrific delay
as it took a lot
of time to re-do.  With silicon design tools, that process is much simpler from
the human's
perspective today, which is a plus.

And let's back up to your first premis.  oxide layers are _part_ of the gate
delay issue.  In
the 80's we had a host of 74xx TTL chips we could choose.  High power.  Low
power.
Shottkey.  You-name-it.  The differences were in the switching times, the power
requirements,
the power dissipation, etc.  That is a part of the silicon design process.  And
it is dictated by the
fab process as I had mentioned...





>>Looking forward to an answer from someone that might carry a little credence
>>in the group here.  :)
>
>I'd really like to see this kind of discussion on RWT (the link given at the
>top), so please post there about it if you're interested in hearing several
>perspectives on it from very knowledgable and experienced people in the field.



This page took 0.04 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.