Transcript lecture16 - Brown University

Design and Implementation of VLSI Systems
(EN0160)
Lecture 16: Scaling Theory
Prof. Sherief Reda
Division of Engineering, Brown University
Spring 2007
[sources: Weste/Addison Wesley – Rabaey/Pearson]
S. Reda EN160 SP’07
Moore’s Law
Moore’s Law. The number of transistors in an integrated circuit
doubles every 2 years.
IBM Cell
234M transistors in
die size of 221 mm2
S. Reda EN160 SP’07
Scaling of MOS transistors
1.2nm
minimum feature size
(gate length)
S. Reda EN160 SP’07
Current oxide thickness ~
1.0 – 2.0nm thickness 
3 – 4 atomic layers of
oxide
Power supply
voltage
scaling
Scaling
Scaling of lithographic wavelength
Fewer and fewer companies can
afford to have their own foundries
S. Reda EN160 SP’07
Avg transistor price is 0.1 μcent!
Number of transistors shipped
Worldwide semiconductor revenues
 Box office: worldwide ~$23B US ($9B); tickets + dvds + tv rights: ~ $44B
 Worldwide total software sales $383B
 Worldwide pharmaceutical sales: $550B
S. Reda EN160 SP’07
Device scaling
(very idealistic NMOS transistor)
(scaled down by L)
scale
doping increased by
a factor of S
Increasing the channel doping density decreases the depletion width
 improves isolation between source and drain during OFF status
 permits distance between the source and drain regions to be scaled
S. Reda EN160 SP’07
Implications of ideal device scaling
S. Reda EN160 SP’07
Total power was increasing (mostly because
of 2 frequency + die) until we hit a “wall”
Intel VP Patrick Gelsinger (ISSCC 2001)
“If scaling continues at present pace, by 2005, high speed
processors would have power density of nuclear reactor, by
2010, a rocket nozzle, and by 2015, surface of sun.”
S. Reda EN160 SP’07
Scaling of standby (leakage) power
Standby power
qV
Poff
1 (  mkTt )

e
tox
bottleneck
 Even if Vt is kept constant after scaling, Poff scales up by S if tox is
scaled down by S
 Vt must be scaled down if VDD is scaled down (otherwise ISAT is
weaker and transistor is slow)
 Standby power would further increase by 10 for every 0.1V
reduction of Vt
S. Reda EN160 SP’07
Power supply voltage (Vdd)
Power/performance tradeoffs
higher
active power
higher
leakage
increasing
performance
Threshold voltage (Vt)
[Taur, 01]
S. Reda EN160 SP’07
Interconnect scaling
w
t
s
h
l
w: width of interconnect (layer dependant)
s: spacing between interconnects with same layer
h: dielectric thickness (spacing between interconnects in two
vertically adjacent layers)
l: length of interconnect
t: thickness of interconnect
S. Reda EN160 SP’07
Constant thickness scaling versus reduced
thickness scaling
reduced thickness scaling
l
w
S
l
w
t
t
t
S
l
S
S. Reda EN160 SP’07
w
constant thickness scaling
w
S
t
l
S
Implications of ideal interconnect scaling
S. Reda EN160 SP’07
Interconnect delay is dominating gate delay
bottleneck
Repeaters can help but…
S. Reda EN160 SP’07
With scaling the reachable radius of a buffer
decreases  we need more and more buffers
bottleneck
repeaters required to
buffer Itanium global
interconnects
 A corner-to-corner (BL-UR) wire in Itanium (180nm) requires 6
repeaters to span die
 Repeaters consume chip area; consume power; add vias
S. Reda EN160 SP’07
Summary
• We are done with the main (and most of
the) material from Chapter 4 (will not
discuss reliability)
S. Reda EN160 SP’07