Intel's 90nm shrink BETTER than their 130nm shrink?

[Tony Hill]

Hi all

I was over at the Ace's Hardware message board, and I saw someone
complaining (again) about how the Prescott P4 chips were using more
power than the Northwood and how they felt that Intel's 90nm shrink
wasn't improving anything.

This got me thinking, just how much did the process actually improve
on a per-transistor basis, so I ran a few numbers and here's what I
got:

Willamette (180nm) 2.0GHz : 42M transistors, 71.8W TDP
Northwood (130nm) 2.0A GHz: 55M transistors, 52.4W TDP

Therefore we get 1.71W/MTrans for Willy and 0.95W/MTrans for
Northwood, or a reduction of 44%

Northwood (130nm) 3.2C GHz: 55M transistors, 82W TDP
Prescott (90nm) 3.2E GHz : 125M transistors, 103W TDP

Here we get 1.49W/MTrans for Northwood vs. 0.82W/MTrans for Prescott,
or a reduction of 45%.


If we were to ignore the cache then things would probably look even
better for the Prescott/Intel's 90nm shrink. Basically every
transistor added in the Willamette -> Northwood shrink were cache
transistors (actually, they only added 13M transistors while an extra
256KB of ECC cache should take up 14M transistors). However with the
Northwood -> Prescott shrink Intel increased the number of non-cache
transistors significantly (from about 25M to 65M transistors as a
rough estimate).

This actually suggests that Intel's 90nm manufacturing process is
doing quite well from a power consumption perspective.

Of course, the flip side to this discussion is that Intel doesn't seem
to have managed to get much performance out of those extra
transistors. Right now they seem to have a chip that is slower, clock
for clock, on most applications when compared to the Northwood. The
theory behind the Prescott is that it will allow for higher total
clock speeds than the Northwood and therefore higher overall
performance. However on the performance/watt scale (one that I've
mentioned as being increasingly important), it seems unlikely that
Prescott will ever match the current Northwood, let alone a 90nm
shrink of a Northwood. Even if Prescott manages to clock to 5.0GHz
while a 90nm shrink of the Northwood would only manage to clock to
4.0GHz it seems likely that, watt for watt, that 90nm "Northwood"
would be the better chip.

Still, when you get right down to it, it's definitely NOT all doom and
gloom for 90nm shrinks either for Intel or anyone else.

 

[Robert Myers]

<snip>

Good post. Interesting observations.

This actually suggests that Intel's 90nm manufacturing process is
doing quite well from a power consumption perspective.

It suggests it, but there are some serious complications in trying to
analyze the situation from the outside (as I'm sure you know). We
don't know what those extra transistors were supposed to do or
actually are doing. Intel isn't going to tell us, and woe to the
Intel employee who leaks anything.

Some of those extra transistors may be doing nothing. That's one way
to get the power consumption/transistor down. Completely baseless
speculation? Yes indeed, but it does fit the known facts.

Long before the actual properties of the process are available fom
testing, Intel starts an expensive core redesign. When the process
becomes available for testing, an unpleasant reality about leakage
emerges.

Clearly, the design has to go on a transistor diet. What to do?
Start over from scratch? Not if you can help it.

Intel has gone through this before with the P4 if I'm not mistaken,
and a prudent company that is capable of learning from its mistakes
might well have designed the chip with some transistors that could be
thrown overboard without a complete redesign. Less performance, less
power, but, oh well, we got the chip out.

Intel has toughed it out before on a new chip release, and they're
going to tough it out again.

As long as I'm in the Prescott speculation department, I might as well
pick up some extras while I'm here. _Do_ look for Prescott to deliver
noticeably better SPEC numbers than its Northwood counterpart. I'll
leave it as an exercise to the reader to figure out how that is going
to happen, given what we already know. ;-).

RM

 

[Tony Hill]

It suggests it, but there are some serious complications in trying to
analyze the situation from the outside (as I'm sure you know). We
don't know what those extra transistors were supposed to do or
actually are doing. Intel isn't going to tell us, and woe to the
Intel employee who leaks anything.

For sure! I'll freely admit that I'm going on relatively little
information here, just some of what is publicly known, which isn't
much!

Some of those extra transistors may be doing nothing. That's one way
to get the power consumption/transistor down. Completely baseless
speculation? Yes indeed, but it does fit the known facts.

There definitely are some parts of the chip that are not enabled
normally. Intel has apparently included some fairly advanced
diagnostic functionality in the chip (though I can no longer find any
reference to this). There are probably other things as well, and I'm
sure that there are more disabled transistors in the Prescott's 125M
than in the Northwood's 55M. However, in the end the ratio of
functioning logic transistors might end up being pretty close to the
same as the ratio of total transistors.

Long before the actual properties of the process are available fom
testing, Intel starts an expensive core redesign. When the process
becomes available for testing, an unpleasant reality about leakage
emerges.

Rumor has it that Prescott consumes almost 40W at idle, suggesting
that leakage current really is a pretty major factor here.

Clearly, the design has to go on a transistor diet. What to do?
Start over from scratch? Not if you can help it.

From my point of view, what you do is enable multi-processor support
and 800MT/s bus on the Pentium-M. These are things that should be
relatively easy to do, it leverages existing technology and it gives
Intel a good performance/watt chip for markets that need it. With 2MB
of L2 cache and 2.0GHz clock speeds this chip should be a heck of a
good server processor except for it's relatively slow (400MT/s) I/O
speeds. It might still be a bit weak for the high-end gamers, but
they usually aren't concerned too much with performance/watt.

Intel has gone through this before with the P4 if I'm not mistaken,
and a prudent company that is capable of learning from its mistakes
might well have designed the chip with some transistors that could be
thrown overboard without a complete redesign. Less performance, less
power, but, oh well, we got the chip out.

It will be interesting to see if they do anything like that with the
"Prescott" Celeron. The current Celeron already demonstrates the
performance is a total non-issue for this line.

Intel has toughed it out before on a new chip release, and they're
going to tough it out again.

As long as I'm in the Prescott speculation department, I might as well
pick up some extras while I'm here. _Do_ look for Prescott to deliver
noticeably better SPEC numbers than its Northwood counterpart. I'll
leave it as an exercise to the reader to figure out how that is going
to happen, given what we already know. ;-).

No need to guess too much, Intel has already published some SPEC
CPU2000 scores on their website:

CINT2000_base
"Northwood" 3.2GHz : 1287
"Prescott" 3.2E GHz : 1363

CFP2000_base
"Northwood" 3.2GHz : 1252 *
"Prescott" 3.2E GHz : 1433

(*) The Northwood CFP results are using version 7.1 of Intel's
compiler, all other results seem to use version 8.0.

The Northwood scores are all from the official SPEC results from
Intel, while the Prescott scores are from Intel's website. There
aren't any details of the system provided on Intel's website though,
so it's probably best that we wait until official results show up
before drawing too many conclusions.

 

[Felger Carbon]

wrote:

<snip>

Good post. Interesting observations.

Here I am agreeing with both Tony and Robert. I must need another
cup of coffee or something like that.

It's becoming evident that Prescott, as an IA32 chip, does not perform
well in smarts/watt. Just the opposite, it's the poster boy for
power inefficiency.

Intel has been focussing (in the x86 world) on maximum clock speed,
since that number is obviously a marketable commodity. IMHO they are
allowing a good x86 implementation, the Pentium M, to languish.

AMD has been focussing on performance, 64-bit capability, and
increasing ASPs at the top end.

Both Intel and AMD have a line of bang/buck CPUs.

But Prescott is a clear signal that the above is insufficient. The
time is closely approaching when smarts/watt is going to be an
important parameter. Intel, almost by accident, is both behind and
ahead in this arena with P4 and P-M.

I agree with everybody else that OoO CPUs perform much better than
in-order CPUs (like Itanium), all other things being equal. But OoO
gives fewer and fewer performance increases at the OoO capability is
increased, while power goes up as OoO capability increases. I predict
that optimum performance/watt will be found when moderate OoO is used.
Netburst, the P4 microarchitecture, is a flat failure as measured by
smarts/watt. P4 is a great success at other things, such as streaming
video.

Will the time come when streaming video folk (and gamesters) will have
to have 10 fans on their PC boxes, while the majority of us will have
at most one fan?

 

[Yousuf Khan]

Felger, you're not a Battlestar Galactica fan are you?

Yousuf Khan

 

[Felger Carbon]

Felger, you're not a Battlestar Galactica fan are you?

Sure am. I'm sure you remember what "felgercarb" was a
network-censor-friendly synonym for! ;-)

Man! The villians in that series, esp. the Cylon honchos, were
characters you could get your teeth into. No namby-pamby naughty
guys. Mean.

 

[Yousuf Khan]

Sure am. I'm sure you remember what "felgercarb" was a
network-censor-friendly synonym for! ;-)

Yeah, exactly why it clicked in me right now. I just watched the original
1978 Battlestar Galactica pilot yesterday and started hearing the word
felgercarb.

Man! The villians in that series, esp. the Cylon honchos, were
characters you could get your teeth into. No namby-pamby naughty
guys. Mean.

Seen the newest incarnation yet?

Yousuf Khan

 

[Keith R. Williams]

@twister01.bloor.is.net.cable.rogers.com>, news.tally.bbbl67
@spamgourmet.com says...

Felger, you're not a Battlestar Galactica fan are you?

Duh! You *just* figured that out? My, my Yousuf, I expected
better of you. ;-)

 

[Yousuf Khan]

@twister01.bloor.is.net.cable.rogers.com>, news.tally.bbbl67
@spamgourmet.com says...

Duh! You *just* figured that out? My, my Yousuf, I expected
better of you. ;-)

I watched Battlestar Galactica so long ago, I didn't even remember it until
I rewatched it recently. Actually all of this time, I had just assumed
Felger was his real name, just one of those strange foreign names like
Yousuf.

Yousuf Khan

 

[Keith R. Williams]

@news04.bloor.is.net.cable.rogers.com>, news.tally.bbbl67
@spamgourmet.com says...

I watched Battlestar Galactica so long ago, I didn't even remember it until
I rewatched it recently. Actually all of this time, I had just assumed
Felger was his real name, just one of those strange foreign names like
Yousuf.

Perhaps I was a little tough on you for not remembering.

....and no, I'm not about to reveal the Felg behind the curtain.
Others have tried to gain access to such secrets of the universe,
but I'll go down defending Truth, Justice, and the American Way!

Besides, the Cylons and Borgs would unite to attack Earth if I
even used FC's initials! ;-)

 

[Yousuf Khan]

Besides, the Cylons and Borgs would unite to attack Earth if I
even used FC's initials! ;-)

You mean to say they aren't already running the Earth?

Yousuf Khan

 

[George Macdonald]

Intel has been focussing (in the x86 world) on maximum clock speed,
since that number is obviously a marketable commodity. IMHO they are
allowing a good x86 implementation, the Pentium M, to languish.

Pentium-M is languishing right enough but is it all intentional? The
1.9GHz CPUs were being talked of months ago but still no sign of them; in
fact they've been stuck at 1.6GHz max for most of that time, with 1.7GHz
just starting to peek its nose out now. If you run those things on heavy
duty work they *do* get bloody hot... I wonder??

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??

 

[V@W]

How come no one has mentioned that prescott is possibly crippled by socket
478 and the current chipsets available?

Maybe mhz for mhz.. a native prescott in LGA775 will be much better than
socket 478 Northwood and socket 478 Precotts?

 

[jack]

: :: Besides, the Cylons and Borgs would unite to attack Earth if I
:: even used FC's initials! ;-)
:
: You mean to say they aren't already running the Earth?

Sure they are. Haven't you noticed G.W. Bush, aka "number 1 of 5" from
the top-echelon Borg packs?

J.

 

[Tony Hill]

Here I am agreeing with both Tony and Robert. I must need another
cup of coffee or something like that.

Cup of coffee?! I think you'll need something a little stronger than
that! :>

It's becoming evident that Prescott, as an IA32 chip, does not perform
well in smarts/watt. Just the opposite, it's the poster boy for
power inefficiency.

Intel has been focussing (in the x86 world) on maximum clock speed,
since that number is obviously a marketable commodity. IMHO they are
allowing a good x86 implementation, the Pentium M, to languish.

I think part of the problem is simply Intel's view of where computing
power is required. The Pentium 4 seems to have been designed to
perform very well for a relatively narrow range of applications (those
that stream lots of data). There are many important applications that
fit this sort of description, but also lots that don't.

What I find particularly interesting is that most server-type
applications don't seem to fit in very well with the P4 mentality.
The Pentium M, on the other hand, does seem to do pretty well as a
server chip and will probably do REALLY well in it's next generation
with 2MB of L2 cache. If Intel were to take this a tiny bit further
and implement SMP support and faster bus speeds, it would probably
make a heck of a server chip. Yet Intel's x86 server line-up remains
firmly routed in the Pentium 4 "NetBurst" architecture. Who needs
streaming video if you're doing transaction processing?!

Perhaps it's some dark, secret plan to push Itanium :>

 

[Tony Hill]

@twister01.bloor.is.net.cable.rogers.com>, news.tally.bbbl67
@spamgourmet.com says...

Duh! You *just* figured that out? My, my Yousuf, I expected
better of you. ;-)

Come now Keith, not everyone is as old as you and Felger! :>

Battlestar Galactica came out the same year that I was born! If it
hadn't been for a bit of a recent resurgence in it (re-releasing on
DVD and doing a new mini-series) I probably wouldn't know anything
about it!

As for Felger's name, I always just figured that it was his porn-name
from back in the 70's. You know: "Long Dong Silver Strikes It Big",
staring Long Dong Silver and Kitty Sinn; Directed by Felger Carbon.
Something like that :>

 

[Tony Hill]

How come no one has mentioned that prescott is possibly crippled by socket
478 and the current chipsets available?

Maybe mhz for mhz.. a native prescott in LGA775 will be much better than
socket 478 Northwood and socket 478 Precotts?

How so? LGA775 doesn't change the bus interface, it just provides a
whole lot of power and grounding pins. It will allow the Prescott to
clock to higher speeds because socket 478 has trouble providing enough
power to the chip, but there are no planned performance enhancements.

Chipsets will improve things a bit, but there's only so much that you
can do there. The i915 and i925 chipsets that will come out alongside
LGA775 add some new features and such, but they aren't expected to
improve performance much. The main goal is to add in DDR2 and PCI
Express support (and remove AGP support). Current rumor is that the
i915 will perform about the same as the current i875, while the i925
will be a tiny bit faster, roughly like the difference between the
i865 and i875 now.

 

[Yousuf Khan]

...and no, I'm not about to reveal the Felg behind the curtain.
Others have tried to gain access to such secrets of the universe,
but I'll go down defending Truth, Justice, and the American Way!

Oh, and BTW, who's behind the Keith Williams curtain? That's way too
obviously a fake name.

Yousuf Khan

 

[Yousuf Khan]

Battlestar Galactica came out the same year that I was born! If it
hadn't been for a bit of a recent resurgence in it (re-releasing on
DVD and doing a new mini-series) I probably wouldn't know anything
about it!

As for Felger's name, I always just figured that it was his porn-name
from back in the 70's. You know: "Long Dong Silver Strikes It Big",
staring Long Dong Silver and Kitty Sinn; Directed by Felger Carbon.
Something like that :>

Well, as it turned out, it was his 70's name. It just wasn't his 70's porn
name, it was his 70's Battlestar Galactica name.

These days, he could go around calling himself Fecal Excrement on the
newsgroups and nobody would even bat an eyelash. Back in the 70's you
had to hide that stuff.

Yousuf Khan

 

[Felger Carbon]

Battlestar Galactica came out the same year that I was born! If it
hadn't been for a bit of a recent resurgence in it (re-releasing on
DVD and doing a new mini-series) I probably wouldn't know anything
about it!

As for Felger's name, I always just figured that it was his porn-name
from back in the 70's. You know: "Long Dong Silver Strikes It Big",
staring Long Dong Silver and Kitty Sinn; Directed by Felger Carbon.
Something like that :>

Alas, Tony, my porn career was tragically cut short by the small
screen format of the day. Had wide-screen HDTV come along two or
three decades earlier, I'd be famous. As it is, I'm only known as a
fanatical supporter of the genius Michael Dell.