2. Why this presentation?
• Everybody loves to talk about Moore’s law starting
from process engineers all the way to singularity
futurists.
• However this talk is different, as it does not have an
extremely gloomy vision of post moore’s law world.
• It is more about connecting different news, data points,
and try to construct a context for them. i.e. trying to
understand the world.
• This presentation will be more handwavy than
scientific (more Michael Moore style than a Gordon
Moore’s one)!!
8. And then the hangover
• First, Power density hit the roof
– Stopped the turbo charging and went back to
dennard scaling to keep power density constant
around 100 W/cm2
• Then, Even Dennard scaling was not possible
since Voltage scaling has slowed down as well,
which means increased power density if
frequency is increased. i.e. frequency scaling
stopped as well.
10. Solutions for single threaded
performance
• Make huge caches! It
improves both
performance and power
efficiency
– 8-12 MB L3 caches in
current high end
processors
• Nehalem Intel® Turbo
Boost Technology: just
yesterday, a laptop turbo
chip (Core i7 mobile) was
introduced by intel.
11. If you can’t improve performance…
• Then we can change what performance means!
• Now we say we are open to parallelism, and
performance is the amount of computation we
get per a certain area and power budget. (i.e. the
Watt/Gflop and mm2/Gflop we get for a certain
power density Watt/mm2 )
• Now dennart scaling improves our newly defined
performance by 8X for halfing the feature size vs
the 2X improvement in single threaded
performance
13. But diminishing returns are catching
up again
• For throughput performance, dennart scaling would
have given 8X improvement for halfing the feature size
with a 2X improvement in single threaded performance
• Because of slowing of voltage scaling we get around 4X
improvement now and we can get another 1.5-2X from
moving to lower energy architectures and no
improvement in single threaded performance
• In the future we will get only 2X improvement when
voltage scaling stops completely and we have to have
degraded single threaded performance
14. From 90 nm to 45 nm is still ok.
Figure 7. Scaling of FMA single and double precision designs from 90nm to
45nm. The performance gain depends on the power density allowed.
15. And future is worse: only 2X factor and
slower individual cores
17. The world has changed as well
“Traditionally, PC makers relied on their software partners to
build bulkier, more demanding applications that required new
hardware. But that strategy has fallen apart. Consumers and
businesses balked at Microsoft’s bloated Vista operating
system, so the company’s upcoming Windows 7 software
should actually require less horsepower and storage space to
run well. The same goes for Apple’s new Snow Leopard
operating system.”
* New York Times, Sep 11, 2009 “Goodbye, Gobbledygook: PC Makers
Abandoning a Sales Pitch Built on Complex Specs”
18. Software people lukewarm reception
of parallelism
“…. I might as well flame a bit about my personal unhappiness with the current
trend toward multicore architecture. To me, it looks more or less like the hardware
designers have run out of ideas, and that they’re trying to pass the blame for the
future demise of Moore’s Law to the software writers by giving us machines that
work faster only on a few key benchmarks! I won’t be surprised at all if the whole
multithreading idea turns out to be a flop, worse than the “Itanium" approach that
was supposed to be so terrific—until it turned out that the wished-for compilers
were basically impossible to write.
Let me put it this way: During the past 50 years, I’ve written well over a thousand
programs, many of which have substantial size. I can’t think of even five of those
programs that would have been enhanced noticeably by parallelism or
multithreading. Surely, for example, multiple processors are no help to TeX.”
Dan E. Knuth, Stanford Professor Emeritus, author of The Art of Computer
Programming and inventor of the TeX computer typesetting system in interview in
2008
19. And the money is not growing fast to
support the more expensive scaling
20. The missing drivers
• Single threaded performance
– At a certain point, a 128 MB cache becomes ridiculous
– As more people get used to stagnating performance, the people who
need performance will move to parallelism
• Throughput performance
– However, a lackluster 2X improvement for halfing the feature size instead
of 8X so far. Worse still, newer machines have to run slower than old
ones.
• Market drivers
– Lack of a killer application and people don’t care. The best sold PCs are
now netbooks
– Software does not push hardware anymore. Think Windows 7 and snow
leopard as leading the software move of the future.
21. So what is next?
• Traditionally, the end of Moore’s law has always been
15 years in the future
• However, this time, we have very high cost of bridging
the frontiers and not matching high returns or high
market demand.
• I would argue that Moore’s law would not be undone
by technical obstacles but rather its irrelevance. The
day we don’t notice the noticeable difference Moore’s
law exerts on our lives is when it will either
dramatically slow down or stop.
• Of course never underestimate the probability of
happy surprises
22. So how does the future of Computing
without Moore’s law look like?
In 1927, GM president Alfred Sloan founded the “Art and Color Section” – the
first design department ever set up by a car manufacturer. GM also had much
success with its strategy of “a car for every purse and purpose”
23. Without moore’s law, we are all
mechanical engineers.
• Design oriented vs Engineering oriented: people
buy cars based on emotional attachment. Same is
happening now with computers. Think Apple.
• Chip manufacturing the Toyota way: check out
the IEEE spectrum article from May 2008 on “the
new economics of Semiconductor
Manufacturing”
24. If you ask the “all-knowing” prescient
market, Moore’s law has stopped already
10000
1000
100
10
1
1975 1980 1985 1990 1995 2000 2005 2010
Nasdaq/100
Computer Performance