Lecture 15: Moore’s Law and
Dennard Scaling
William Gropp
[Link]/~wgropp
What Do You Think Moore’s
Law Says?
• Usually cast as X doubles every
18-24 months.
• Is X:
♦ Computer performance
♦ CPU Clock speed
♦ The number of transistors per chip
♦ One of the above, at constant cost?
2
The Original Moore’s Law
• The number of transistors per chip,
at constant cost
• This has not been true for years
• The improvement has been
remarkable, but it is getting
increasingly difficult to maintain
this exponential improvement
3
Road Maps
• The Semiconductor industry has
produced a roadmap of future trends
and requirements
♦ Semiconductor Industry Association
(~1977, roadmaps from early ’90s)
♦ International Technology Roadmap for
Seminconductors (~1998)
• ITRS Home and Current Summary
♦ [Link]
♦ [Link]
2013TableSummaries/
2013ORTC_SummaryTable.pdf
4
ITRS projections for gate lengths (nm)
for 2005, 2008 and 2011 editions
30
Physical gate length, nm
25
3 years
20 ITRS
edition
5 years
15
2005
10 2008
2011
5 2013
0
2007 2009 2011 2013 2015 2017 2019 2021 2023 2025 2027
Expected manufacturing year
Note the rapid 3- and then 5-year shifts in ITRS projections for
physical gate lengths.
5
Why Is Moore’s Law
Confused with Performance?
• Its not because of parallelism – for
most of its life, Moore’s Law and
single processor performance
correlated well
• The reason is the size and speed
are related – the smaller
something is, the quicker it can be
changed
♦ Thus, smaller transistors can switch
at higher speeds
6
Clock Speed History
7
Dennard Scaling
• Why haven’t clock speeds increased,
even though transistors have continued
to shrink?
• Dennard (1974) observed that voltage
and current should be proportional to
the linear dimensions of a transistor
♦ Thus, as transistors shrank, so did
necessary voltage and current; power is
proportional to the area of the transistor.
8
Dennard Scaling
• Power = alpha * CFV2
♦ Alpha – percent time switched
♦ C = capacitance
♦ F = frequency
♦ V = voltage
• Capacitance is related to area
♦ So, as the size of the transistors
shrunk, and the voltage was reduced,
circuits could operate at higher
frequencies at the same power
9
End of Dennard Scaling
• Dennard scaling ignored the
“leakage current” and “threshold
voltage”, which establish a
baseline of power per transistor.
♦ As transistors get smaller, power
density increases because these don’t
scale with size
♦ These created a “Power Wall” that
has limited practical processor
frequency to around 4 GHz since
2006 10
Historical Clock Rates
10,000
1,000
Clock (MHz)
100
10
1/1/80
1/1/84
1/1/88
1/1/92
1/1/96
1/1/00
1/1/04
1/1/08
1/1/12
Figure 3.13. Historical Clock Rates.
11
3.3.3 Core Clocks
Prediction Trends
100
On Chip Clock Rate (GHz)
10
1
1995 2000 2005 2010 2015 2020 2025
1999 2001 2004
2006 2008 2010
2011 2012 TOP500!top!clock
Figure 3.14. On-chip Clock from ITRS.
12
conservative projection assuming very heavily power-limited designs, with a CAGR of a mere
4.3%. At least for 2012 these estimates appear to reflect reality.
Questions
• True or False: Moore's law says
that computer performance
doubles every 18-24 months
• True or False: Moore’s law allows
us to predict future properties in
the same way that the law of
gravity allows us to predict the
path of a planet around the sun.
13
Questions
• If the number of transistors per
unit area (i.e., on the same chip
size) doubles every 24 months:
♦ By what factor does the length of the
side of a transistor, assuming it is
square, decrease every year?
♦ Starting with a length of 18.4nm
(2014), what would be the size of a
transistor in 2025? How does this
compare to the ITRS value of
6.75um?
14
