[Beowulf] HPC Market Question

Mark Hahn hahn at mcmaster.ca
Mon Jan 19 07:25:50 PST 2009


> It means that Core2 Quad were architected as two dual-core chips sitting on
> the same silicon die.

well, probably not - using two separate die would improve yield,
since each would be independently testable.  but this is irrelevant:
moore's law is about the process-driven t-count for cost-effective chips.
the actual chips that vendors choose to produce is dominated by other
factors.  consider that Intel produced much larger ia64 chips before 
the modest-sized i7.  or for that matter _did_ produce 6-core dunnington
based on core2 and 45nm.  or that Nvidia's high-end chip is similarly large.

the real point here is that moore's law is mainly about transistor count
is an area process, driven by moderately steady progress in reducing litho
dimensions.  there have been bobbles in litho before, and will almost 
certainly happen around 22-32nm as well.  but shrink delays are to some 
extent offset by design changes (multicore uses up lots of space without
a lot of design effort) or even transistor-level changes (I'm thinking of
multi-gate, rather than hi-k/metalgate).

one interesting thing is clock frequency - a colleague of mine loves the idea
that the "clock scaling" era is over, but I disagree.  there are various
design trends - themes really, but they're not terminal.  I see no reason to
expect that clocks will always remain in the 2-3 GHz range in the future,
especially once the multicore trend is "played out" (ref recent unsurprising
Sandia press.)  yes, there surely is a tradeoff between clock and power given
a fixed process technology.  but it's not fixed, and besides, why would the 
market continue to reward improvements in power?  65W quad-core cpus seem 
pretty nice to me in the desktop environment, especially with good clock/core
modulation and intelligent design of other components.  is it that laptops
are now driving the market?



More information about the Beowulf mailing list