Processor Speed

[Charles Turner]

Hi,

Sorry if this is a bit of a basic question but my understanding is that the
smaller you can make the devices the faster potentially they can perform.
However there is a limit. Considering the speed of processors in PC's has
increased from 100MHz or so to 3GHz, what is the fastest speed that will be
attainable?

 

[Rob Stow]

Hi,

Sorry if this is a bit of a basic question but my understanding is that the
smaller you can make the devices the faster potentially they can perform.
However there is a limit. Considering the speed of processors in PC's has
increased from 100MHz or so to 3GHz,

I've used PCs with x86 processors for 18 or 19 years and clock
speeds have gone up from 4 MHz to 3.6 GHz in that interval.

what is the fastest speed that will be
attainable?

I read something a few months ago about transistors running at
30 GHz in a lab. As to what we are likely to see outside of a
lab in the foreseeable future ???

And clock speeds aren't everything - I'll take a 2.4 GHz AMD64
chip over a 3.6 GHz Intel chip any day of the week.

 

[Yousuf Khan]

Hi,

Sorry if this is a bit of a basic question but my understanding is that the
smaller you can make the devices the faster potentially they can perform.
However there is a limit. Considering the speed of processors in PC's has
increased from 100MHz or so to 3GHz, what is the fastest speed that will be
attainable?

Well, that used to be the case upto and including the 130nm node, but
it's not the case as much anymore. There used to be a time when you
could take the same processor design, shrink it down by half, and you'd
also automatically be able raise clock speeds by double or more. That
was simply because the shrinkage would let the electrons travel shorter
distances. However, now due to quantum mechanical effects, the electrons
have a tendency to jump right out of their conductors, and that makes
them slower.

Yousuf Khan

 

[George Macdonald]

Hi,

Sorry if this is a bit of a basic question but my understanding is that the
smaller you can make the devices the faster potentially they can perform.
However there is a limit. Considering the speed of processors in PC's has
increased from 100MHz or so to 3GHz, what is the fastest speed that will be
attainable?

Well Intel has completely backed away from their "expectation" of 10GHz by
the end of the decade and has taken a path which leads (back ?) to a CPU
which does more work per clock cycle. In fact they balked on the P4 at
4GHz... never happened at 90nm.

Here's an interview with Meyerson of IBM where he outlines the problems of
power density and leakage with recent 90nm geometry shrinks which have led
to the end of "classical scaling":
http://www.reed-electronics.com/electronicnews/article/CA508575 and another
article on an IBM scientist who fills in the picture on the new importance
of materials:
http://www.reed-electronics.com/electronicnews/article/CA6281310

When you get down to geometry features in semiconductors which are a few
atoms thick, the laws of physics get in the way. Result: dual core CPUs.

 

[nobody]

Hi,

Sorry if this is a bit of a basic question but my understanding is that the
smaller you can make the devices the faster potentially they can perform.
However there is a limit. Considering the speed of processors in PC's has
increased from 100MHz or so to 3GHz, what is the fastest speed that will be
attainable?

No matter what it is, probably it will never be attained because it
will require certain compromises (waaaay longer pipelines, to name
just one) that hurt the performance more than you gain from raw Giga
hurts increase ;P
After all, what's the goal? Get into Guinness book with the top
number, or get the most performance from the existing technology,
preferably for less expenses? Even Intel finally got the
understanding, even though they preached the contrary for long, long
time.

NNN

 

[Rob Stow]

No matter what it is, probably it will never be attained because it
will require certain compromises (waaaay longer pipelines, to name
just one)

Intel thought they needed super long pipelines for the
architecture used by the P4s and the P4-Xeons. However, you
can't assume from that one rotten example of CPU architecture
that every CPU design will also require long pipelines in order
to reach the same clock speeds.

AMD, for example, has hit 2.8 GHz - twice the speed of the
initial 1.4 GHz P4's - with the AMD64 architecture and I have
seen an AthlonFX overclocked to 3.4 GHz using air-cooling.

 

[Tony Hill]

Intel thought they needed super long pipelines for the
architecture used by the P4s and the P4-Xeons. However, you
can't assume from that one rotten example of CPU architecture
that every CPU design will also require long pipelines in order
to reach the same clock speeds.

AMD, for example, has hit 2.8 GHz - twice the speed of the
initial 1.4 GHz P4's - with the AMD64 architecture and I have
seen an AthlonFX overclocked to 3.4 GHz using air-cooling.

I'm not sure that AMD is all that better off here than Intel. Their
Athlon64 FX chips from their 130nm production lines topped out at
2.6GHz. Their 90nm production line has thus far only managed to get
things up to 2.8GHz. I figure they'll probably clear 3.0GHz before
they're done, maybe up to 3.2GHz in about a years time, but not much
more out of 90nm. That's only a 23% increase in clock speeds,

 

[Keith]

When you get down to geometry features in semiconductors which are a few
atoms thick, the laws of physics get in the way. Result: dual core CPUs.

Actually, I think dual-cores came about because there wasn't
anything else interesting to do with transistors. Large caches were
the first useful transistor-sink but sooner or later, diminishing
returns bites hard. The next obvious place to absorb transistors
was another core.

At least that was my proposal here, oh, about the turn of the
century. ;-)

 

[Rob Stow]

Actually, I think dual-cores came about because there wasn't
anything else interesting to do with transistors. Large caches were
the first useful transistor-sink but sooner or later, diminishing
returns bites hard.

Within an unlimited transistor budget I would be dreaming about
several gigs of on-chip RAM - skip the L2 and off-chip RAM
completely. I have images of a chip so big that you put in on
the /other/ side of the motherboard ...

And I think of the "other side of the motherboard" idea in the
context of "ordinary" CPUs as well. Imagine if the left side of
a tower or mid-tower case was a single, huge passive heat-sink
and your motherboard was attached directly to it - no intervening
back plane - with the processor(s) firmly clamped between the
motherboard and the case-panel/heatsink.

 

[The little lost angel]

And I think of the "other side of the motherboard" idea in the
context of "ordinary" CPUs as well. Imagine if the left side of
a tower or mid-tower case was a single, huge passive heat-sink
and your motherboard was attached directly to it - no intervening
back plane - with the processor(s) firmly clamped between the
motherboard and the case-panel/heatsink.

Something like the Zalman heatsink that was a casing I saw somewhere
before? Heard it was a major PITA to install/uninstall anything

 

[Keith]

Within an unlimited transistor budget I would be dreaming about
several gigs of on-chip RAM - skip the L2 and off-chip RAM
completely. I have images of a chip so big that you put in on
the /other/ side of the motherboard ...

Two obvious problems here. We're not talking about infinite numbers of
transistors, only nearly so. ;-) DRAM also is still a costly process to
add to the processor recipe.

And I think of the "other side of the motherboard" idea in the
context of "ordinary" CPUs as well. Imagine if the left side of
a tower or mid-tower case was a single, huge passive heat-sink
and your motherboard was attached directly to it - no intervening
back plane - with the processor(s) firmly clamped between the
motherboard and the case-panel/heatsink.

Sounds like a tollerance problem here. You really don't want any strain
on the MB. The other problem is that even with an infinitely massive
heatsink the power density isn't infinite. The heat still has to get off
the chip.

 

[Rob Stow]

Two obvious problems here. We're not talking about infinite numbers of
transistors, only nearly so. ;-)

"Dream the impossible dream." Beer helps.

DRAM also is still a costly process to
add to the processor recipe.


Sounds like a tollerance problem here. You really don't want any strain
on the MB.

Why do you think there would be any more strain on the
motherboard than one currently gets when mounting the motherboard
on a much less sturdy backplane ?

In case it is not clear, what I imagine would not be a case of
mounting a heatsink on a motherboard - it would be more akin to
mounting a motherboard on a heatsink.

The other problem is that even with an infinitely massive
heatsink the power density isn't infinite. The heat still has to get off
the chip.

I would imagine that pipes embedded within the sink would help.
As to whether it would help enough ???

 

[Ed]

I'm not sure that AMD is all that better off here than Intel. Their
Athlon64 FX chips from their 130nm production lines topped out at
2.6GHz. Their 90nm production line has thus far only managed to get
things up to 2.8GHz. I figure they'll probably clear 3.0GHz before
they're done, maybe up to 3.2GHz in about a years time, but not much
more out of 90nm. That's only a 23% increase in clock speeds,

But with multi-core chips that would be 23% boost per core. ;p
Ed

 

[Keith]

"Dream the impossible dream." Beer helps.

Enough beer and you don't dream at all. ;-)

Why do you think there would be any more strain on the
motherboard than one currently gets when mounting the motherboard
on a much less sturdy backplane ?

If I understand your proposal, you're attaching the processor to the
case. The motherboard is attached to the case. The
processor/socket/spreader are sandwitched inbetween. If things aren't
perfect you're not going to have decent contact with the case (processor
stack too skinny) or the board is going to be bowed (too thick). Either
way there is stress on the board.

In case it is not clear, what I imagine would not be a case of mounting
a heatsink on a motherboard - it would be more akin to mounting a
motherboard on a heatsink.

Think about shipping the thing by USPS. That's a lot of torque on the
processor stack/board.

I would imagine that pipes embedded within the sink would help. As to
whether it would help enough ???

Even heat pipes aren't an infinite sink. You still have to get the power
off the chip. Even the silicon doesn't have zero thermal resistivity.

 

[Tony Hill]

But with multi-core chips that would be 23% boost per core. ;p

Err, AMD's fastest dual-core chip runs at 2.4GHz, with a 2.6GHz core
expected soon. They'll probably remain at about 2 speed grades slower
than their single-core counterparts, so we're looking at topping out
at (likely) 2.8GHz for dual-core chips.

That's only an 8% increase per core vs. the 130nm parts.

The story isn't all that different at Intel either, with dual-core
90nm chips maxing out at 3.2GHz now, maybe reaching 3.4GHz before
their through. This is exactly the same clock speeds as for dual-core
90nm chips as where the single-core 130nm chips finished their
lifespan.

 

[Del Cecchi]

Actually, I think dual-cores came about because there wasn't
anything else interesting to do with transistors. Large caches were
the first useful transistor-sink but sooner or later, diminishing
returns bites hard. The next obvious place to absorb transistors
was another core.

At least that was my proposal here, oh, about the turn of the
century. ;-)

That was not the reason exactly on Power4

 

[Keith]

That was not the reason exactly on Power4

Actually it was, pretty much. What else were they going to do with
the transistors? Of course they also added a few more coppies of the chip
and an L3 on the same MCM.

Also note the 970 -> 970MP (which is essentially half a Power4,
timing tweaked, with VMX and a different bus -> doubled .

 

[Greg]

Hi,

Sorry if this is a bit of a basic question but my understanding is that the
smaller you can make the devices the faster potentially they can perform.
However there is a limit. Considering the speed of processors in PC's has
increased from 100MHz or so to 3GHz, what is the fastest speed that will be
attainable?

There are many limiters for a chips speed, from raw transistor size up
to macro-scopic packaging.

At a packaging and manufacturer level, the CPU is spec'd to dissapate a
given themal power. It is easy to get faster speed by simply running
smaller transistor dimensions, but the off-state leakage exponentially
increases for only a linear increase in speed.

There are certainly others here who can comment more knowledgeably on
the design issues, but there are significant timing corrections to sync
signals that were more even more troublesome on Prescott than
Northwood.

At a transistor level, there are several technical challenges to
overcome. Typically, the gate oxide and transistor dimensions are
scaled with each technology generation to improve short-channel effects
and increase drive current. However, the SiON gate oxide is near its
limit and can't be scaled further without degrading oxide reliability.
The transistor dimension can be scaled down further (90nm -> 65nm
technology) which will decrease the parasitic gate-source/drain
capacitance and reduce transit time of carriers across the channel.
However, without gate oxide scaling these devices will suffer from
degraded short-channel effects.

-Greg

 

[Del Cecchi]

Actually it was, pretty much. What else were they going to do with
the transistors? Of course they also added a few more coppies of the chip
and an L3 on the same MCM.

Also note the 970 -> 970MP (which is essentially half a Power4,
timing tweaked, with VMX and a different bus -> doubled .

My recollection was that they needed a chip bad, what with the
outstanding success of bellatrix and habanero and those. So it was
decided to do a straightforward core, and put two of them on, rather
than a more complex core. That made it more likely that the chip would
actually work. As I recall, they wound up cutting the size of the cache
rather than dumping a core when the chip size turned out to be bigger
than the foils said.

 

[Keith]

My recollection was that they needed a chip bad, what with the
outstanding success of bellatrix and habanero and those. So it was
decided to do a straightforward core, and put two of them on, rather
than a more complex core. That made it more likely that the chip would
actually work. As I recall, they wound up cutting the size of the cache
rather than dumping a core when the chip size turned out to be bigger
than the foils said.

Maybe. I wasn't in on the Power4's politics. However, it surely
didn't end up as a "straightforward core". Five instructions
issued/completed per cycle and a couple of hundred instructions in
flight isn't exactly "straightforward". Sure, more cache would
have been nice, but the second core was nicer. Apparently they
traded off "straightforward" and larger caches for better
performance. Ok, I'll accept that that's not the same as adding
another core because it "was something to do with transistors". I
maintain that they would have improved single-core perforamnce with
those transistors (apparently they did that too), were that
possible.