Yeah, more pipelining, adding more dispatch slots, rename registers and
an industry leading out-of-order window is really baaaaad ... NOT! The
I think that's the point, most of those features were stripped out of
the cores for both the PPC core of the Cell processor for PS3 and the
PPC chips for XBox360. It remains to be seen if the G5 itself can
clock to 3.2GHz on a 90nm process, or perhaps even on a 65nm process.
See the clock scaling business is clearly something AMD learned with
the K6, and Intel learned with the original Pentium.
The basic concept of clock scaling has been something that has been
fairly well known for some time. For a good example of this, look all
the way back to '92 and DEC's Alpha processor. It's always been a
trade-off between clock scaling, instructions per clock cycle, power
consumption and the all-important factor of cost.
Just looking at
the history of these two companies, and you can see that both know how
scale a processors clock throughout its lifetime to match up with
Moore's Law. The K6 went from 233Mhz to 550Mhz, the Pentium 33Mhz to
200Mhz, Athlon from 600Mhz to 2.0Ghz, Pentium II/III from 300Mhz to
1.266Ghz, Pentium IV from 1.5Ghz to 3.8Ghz, and the Athlon-64 got
started at around 2.0Ghz. Look at the anemic PPC 970 processor;
introduced at 2.0 Ghz, now a measley 2.5 Ghz but only by using *liquid
cooling*.
Well, to be fair to IBM, their original design goal for the PPC 970/G5
was 1.6 and 1.8GHz. They ended up getting an extra speed grade of
2.0GHz to sell to Apple. Now they are up to 2.7GHz with the latest
and greatest PowerMac chips. It may not seem like great scaling, but
pretty much everyone else has had similar (if not worse) woes in the
past while. Intel hit 3.2GHz a full 2 years ago, yet their fastest
clocked chip now is still only 3.8GHz.
As for the liquid cooling this, Apple definitely *COULD* cool the
chips using air if they so desired, but it would require rather noisy
fans to do so, especially with two processors in the box.
And lets talk about power consumption. Because of the threat from
Transmeta (with their amazing low power cores) both AMD and Intel have
reacted by making lower and lower power cores.
I don't think Transmeta had much to do with this, it's more a question
that the market WANTED lower power cores. Intel has been producing
extremely low powered cores for some time now, certainly dating back
to before Transmeta showed up on the scene, they just started getting
more popular and Intel saw it as a business opportunity.
It's also important to note that a lot of these "low power" cores are
actually consuming quite a lot of power relative to what chips used to
consume. For example, the Pentium-M consumes about 25W. An old
PentiumMMX 300MHz Mobile processor consumed only 8.4W. Even the
PentiumMMX 233MHz DESKTOP chip only consume 17.0W of power.
For comparison, Transmeta's "amazing low power cores" consume a
maximum of about 7 or 8W (exact numbers are very poorly documented by
Transmeta).
Intel will eventually
switch to them in their mainstream processors (via the Pentium-M path)
and AMD has put their low-power technology (gates clocks, and other
transistor techniques) into their entire product-line. In the
meantime, IBM, not feeling any competitive pressure from anyone, just
decided to crank the power consumption through the roof.
IBM is another company that does a very poor job of (publicly)
documenting power consumption, however my understanding is that the
PowerPC 970FX tops out at less than 100W. This puts it well into the
same ballpark as the Athlon64 or P4.
[...] and Sony is willing to accept at 12.5% defect rate (1 SPE out of 8
per die being nonfunctional).
Interesting. I am not aware of any other CPU manufacturer willing to
share their yeild rates publically, so I don't really know what to
compare that to.
I don't think that this is exactly the same thing as yield rate, more
just that allowing a certain amount of redundancy allows for increased
yields. AMD and Intel (and all others) do this with cache, where a
processor with 1MB of cache is actually built using more than 1MB and
then disabling any defective blocks. The design of The Cell just
allows IBM to extend this to processing units in addition to cache.
FWIW though, typically anything over 80% is considered good yields for
a high-end processor.
That's right IBM, cannot hope to accomplish what even the lowly humble
AMD has done. IBM simply doesn't have the competence, drive, or will
to compete with either Intel or AMD.
I think it's as much a question of return on investment as anything
else. Even with selling ~35M processors a year AMD is only just
breaking even. With IBM selling only about 5M processors to Apple
there's no way that they would have been making much money, if any at
all. Certainly their R&D costs would have been partially covered by
other processor sales (eg. the core of the PPC 970 is partly based off
of the Power4 processor), but I can't see it as being a particularly
good business venture for them.
With the console businesses IBM should be generating at least 50-75M
unit sales per year. Even with much lower profit margins it is likely
that this business will be a lot more profitable.
