Dual Core - Is this a battle AMD can win?

[Greg]

I got thinking about the impact to dual-core technology and whether
AMD has painted itself into a corner.

AMD wins points for the "coolness" factor of announcing the first x86
dual-core. They already won points for 64bit, not that it is useable
outside linux at this point. BUT... I wonder if this is the best
approach.

My concerns with dual-core:
1) Two CPU's aren't as efficient with the resources as one. Dualie
systems are only ~60% better performing over a single core.
Admittedly dual-core systems are different beasts here, but it is
certainly not going to be 2x performance.

2) Huge impact on the die per wafer yielded. In addition to fewer
available units, any defect that kills one of the dual CPUs will kill
the unit as a dual-core.

It is really (2) that bothers me, especially since we are talking
about going to mainstream desktop dual-core systems. Although AMD has
partnered with IBM and Chartered, the overall 300mm capacity for AMD
is significantly less than Intel which has 4 300mm fabs online now and
another in retro. Intel can bury AMD in silicon.

AMD has had better CPU designs, which has allowed it to gain
marketshare in mainstream and server space. I wonder why AMD did not
leverage their design teams to engineer a better solution than a
seeming "desperate" switch to dual-core. Doesn't this move play into
Intel's capacity advantage?

-Greg

 

[Never anonymous Bud]

Trying to steal the thunder from Arnold, (e-mail address removed) (Greg) on 19 Nov 2004 12:37:26 -0800 spoke:

My concerns with dual-core:
1) Two CPU's aren't as efficient with the resources as one. Dualie
systems are only ~60% better performing over a single core.
Admittedly dual-core systems are different beasts here, but it is
certainly not going to be 2x performance.

But you WILL be able to run some software faster, even if only
60%, and a dual-CPU system is cheaper AND smaller than 2 single-CPU
systems, which WON'T work the same as a dual-CPU system (on a single
large graphic file, ferinstance).

Intel can bury AMD in silicon.

Yeah, like THAT matters.

 

[George Macdonald]

I got thinking about the impact to dual-core technology and whether
AMD has painted itself into a corner.

AMD wins points for the "coolness" factor of announcing the first x86
dual-core. They already won points for 64bit, not that it is useable
outside linux at this point. BUT... I wonder if this is the best
approach.

My concerns with dual-core:
1) Two CPU's aren't as efficient with the resources as one. Dualie
systems are only ~60% better performing over a single core.
Admittedly dual-core systems are different beasts here, but it is
certainly not going to be 2x performance.

Depends on the task(s) of course but if that 60% is an "average" figure,
expect it to improve. For mainstream computing, programmers will adapt to
make gains and there are already numerically intensive & DP tasks which
yield >100% efficiency on dual and higher multiple cores. Besides that,
there is the system response you get with dual, which *is* appreciated by
even the desktop market.

2) Huge impact on the die per wafer yielded. In addition to fewer
available units, any defect that kills one of the dual CPUs will kill
the unit as a dual-core.

I have to assume the "model" has been studied in detail and it "works" -
there'll always be a low-end Celeron/Duron-Sempron.

It is really (2) that bothers me, especially since we are talking
about going to mainstream desktop dual-core systems. Although AMD has
partnered with IBM and Chartered, the overall 300mm capacity for AMD
is significantly less than Intel which has 4 300mm fabs online now and
another in retro. Intel can bury AMD in silicon.

By the time we get to duals, AMD will have their exisitng Dresden, plus the
new 300mm there, as well as the foundry deals. Even if they're targeting
say 30-40% market share, I'd think they will have sufficient capacity.

AMD has had better CPU designs, which has allowed it to gain
marketshare in mainstream and server space. I wonder why AMD did not
leverage their design teams to engineer a better solution than a
seeming "desperate" switch to dual-core. Doesn't this move play into
Intel's capacity advantage?

If you believe the talk of "the end of scaling", and that is the line from
some top industry insiders, short of a major leap in technology, where else
is there to go but dual cores? I'm not sure that the K8 is really that
much better a CPU "design" than the Pentium-M; Intel just went off on its
P4 adventure and it looks like it will now self-correct.

Intel is getting some mileage out of Hyper Threading and I haven't heard
talk of AMD following so, in a sense, dual is AMD's answer there.

Rgds, George Macdonald

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

 

[Tony Hill]

I got thinking about the impact to dual-core technology and whether
AMD has painted itself into a corner.

AMD wins points for the "coolness" factor of announcing the first x86
dual-core. They already won points for 64bit, not that it is useable
outside linux at this point.

And *BSD and Solaris... Basically everyone other than Microsoft, who
were supposed to have their 64-bit OS out over a year ago, but like
everything else Microsoft does, it's horribly late.

BUT... I wonder if this is the best
approach.

My concerns with dual-core:
1) Two CPU's aren't as efficient with the resources as one. Dualie
systems are only ~60% better performing over a single core.
Admittedly dual-core systems are different beasts here, but it is
certainly not going to be 2x performance.

Sure, but what the heck else are you going to do with all the die
space?!

2) Huge impact on the die per wafer yielded. In addition to fewer
available units, any defect that kills one of the dual CPUs will kill
the unit as a dual-core.

Sure, but again, what the heck else are you going to do with all that
die space?

Take the example of Intel's Northwood vs. Prescott. Intel more than
doubled the number of transistors. If the two chips were built on an
equivalent process, the Prescott would be SIGNIFICANTLY larger than
the Northwood, and yet in the end it resulted in almost non-existent
performance gains. Why? Because we're rapidly hitting a point of
diminishing returns for these chips. Even if you throw a TON more
execution units and decoders and various other widgets, you tend to
only get a pretty small increase in performance. We've apparently
reached the stage now where doubling these things results in very
small improvements in performance. The main problem here being that
memory latency has skyrocketed relative to a processors raw number
crunching ability.

There are a few solutions here. First and most simple is adding more
cache, but that only gets you so far and eventually wears out as well.
The second is do to some sort of multithreading. Good for some
situations, but it bites you in the ass on others. The final solution
is to go dual-core.

It is really (2) that bothers me, especially since we are talking
about going to mainstream desktop dual-core systems. Although AMD has
partnered with IBM and Chartered, the overall 300mm capacity for AMD
is significantly less than Intel which has 4 300mm fabs online now and
another in retro. Intel can bury AMD in silicon.

Going to dual-cores only roughly doubles the number of transistors
used in a chip, and doubling the number of transistors used in a chip
every 18-24 months has been standard-fare for the past 30 years or so
(Moore's law and all). Designers simply have to look forward and say
"what is the best way that we can spend the number 2x increase in
transistors". There are lots of options, but I think pretty much all
designers have come to the conclusion that, at this stage at least,
going to dual-cores is the best option.

Remember, dual-core 90nm chips are roughly the same size as
single-core 130nm chips. Same goes for 65nm vs 90nm.

AMD has had better CPU designs, which has allowed it to gain
marketshare in mainstream and server space. I wonder why AMD did not
leverage their design teams to engineer a better solution than a
seeming "desperate" switch to dual-core. Doesn't this move play into
Intel's capacity advantage?

One could easily argue that it is just the opposite. AMD increased
the number of transistors in the Opteron nearly 3-fold from the
AthlonXP (Thoroughbred at least). How much performance did that gain
them? How much additional performance would an additional 2-fold
increase in performance given that they've already used up their
integrated-memory-controller card? An extra 10%? 5%? Less?

If AMD wanted to go head-to-head with Intel in single-core
performance, Intel could potentially just throw a LOT more die space
(ie for cache) at their solution. Case-in-point, the new P4's that
have 2MB of L2 cache (first one just appeared as an official Intel
part earlier this week).

Relative to bumping up single-core performance, going dual-core is
pretty cheap (die-space wise) way of gaining performance.

 

[chrisv]

I got thinking about the impact to dual-core technology and whether
AMD has painted itself into a corner.

I don't see how could have "painted itself into a corner". It's not
like they're going to stop producing single-core CPU's.

AMD wins points for the "coolness" factor of announcing the first x86
dual-core.

Announcements aren't cool. Show me the silicon.

(snip)

AMD has had better CPU designs, which has allowed it to gain
marketshare in mainstream and server space. I wonder why AMD did not
leverage their design teams to engineer a better solution than a
seeming "desperate" switch to dual-core. Doesn't this move play into
Intel's capacity advantage?

It's hardly "desperate". It's more like "an obvious thing to do, at
the high end". My prediction is that dual-cores will be marketed as a
very expensive, high-end solution, out to a few years from now. This
will solve all of the "yield" issues that you raised - they'll just
charge more to compensate.

 

[keith]

I got thinking about the impact to dual-core technology and whether
AMD has painted itself into a corner.

AMD wins points for the "coolness" factor of announcing the first x86
dual-core. They already won points for 64bit, not that it is useable
outside linux at this point. BUT... I wonder if this is the best
approach.

My concerns with dual-core:
1) Two CPU's aren't as efficient with the resources as one. Dualie
systems are only ~60% better performing over a single core.
Admittedly dual-core systems are different beasts here, but it is
certainly not going to be 2x performance.

SMP is better at running forground tasks (interactive work or latency from
the user's perspective, depending on how you look at it) than your 60%
would indicate.

2) Huge impact on the die per wafer yielded. In addition to fewer
available units, any defect that kills one of the dual CPUs will kill
the unit as a dual-core.

What makes you think that a defect in one makes the other useless? It's
not quite like a 2X (or 4x) hit anyway, since much of the area is in
arrays where redundancy is often used to improve productivity.

It is really (2) that bothers me, especially since we are talking about
going to mainstream desktop dual-core systems. Although AMD has
partnered with IBM and Chartered, the overall 300mm capacity for AMD is
significantly less than Intel which has 4 300mm fabs online now and
another in retro. Intel can bury AMD in silicon.

....only if they have a sucker to buy it. ;-)

AMD has had better CPU designs, which has allowed it to gain marketshare
in mainstream and server space. I wonder why AMD did not leverage their
design teams to engineer a better solution than a seeming "desperate"
switch to dual-core. Doesn't this move play into Intel's capacity
advantage?

I don't see it in any way "desperate". Everyone is going to SMP. It's a
natural, given limited design budget and "unlimited" transistor budget. I
don't see how this plays into any "capacity advantage" at all. A 90nm MP
should be about the same size as a 130nm UP.

 

[ykhan]

1) Two CPU's aren't as efficient with the resources as one. Dualie
systems are only ~60% better performing over a single core.
Admittedly dual-core systems are different beasts here, but it is
certainly not going to be 2x performance.

AMD's existing multiprocessors are getting around 80% scaling
efficiency according to some HPC tests. And it's scaling efficiency is
maintained from 2 processors all of the way upto its maximum limit of
8 processors; it doesn't drop off in efficiency the more processors
that you add. That's mainly due to the point-to-point Hypertransport
bus.

A dual-core should scale up with even more efficiency since the
internal short-distance HT links are going to be much faster than the
external long-distance HT links.

2) Huge impact on the die per wafer yielded. In addition to fewer
available units, any defect that kills one of the dual CPUs will kill
the unit as a dual-core.

It might kill it as a dual-core true, but half of it might still be
salvageable as a single-core chip. They are using the same socket
whether it's single- or dual-core.

It is really (2) that bothers me, especially since we are talking
about going to mainstream desktop dual-core systems. Although AMD has
partnered with IBM and Chartered, the overall 300mm capacity for AMD
is significantly less than Intel which has 4 300mm fabs online now and
another in retro. Intel can bury AMD in silicon.

AMD has had better CPU designs, which has allowed it to gain
marketshare in mainstream and server space. I wonder why AMD did not
leverage their design teams to engineer a better solution than a
seeming "desperate" switch to dual-core. Doesn't this move play into
Intel's capacity advantage?

I'm sure Intel is hoping so. In fact, I've even speculated that Intel
might just try to produce a large percentage of dual-cores and just
relabel failed dualies as singulars. It's got five 300mm plants, and
it's got an overcapacity problem, so it's got lots of factories that
it needs to keep busy.

On AMD's side, it has the advantage of designing for dual-cores well
ahead of time, therefore it now has enough time to tweak its
manufacturing process until it is really efficient. So that it doesn't
need to waste any chips in a brute force fashion.

Yousuf Khan