Transcript ESE680-002 (ESE534): Computer Organization Day 6: January 29, 2007 VLSI Scaling Penn ESE680-002 Spring 2007 -- DeHon.

ESE680-002 (ESE534):
Computer Organization
Day 6: January 29, 2007
VLSI Scaling
1
Penn ESE680-002 Spring 2007 -- DeHon
Today
•
•
•
•
•
VLSI Scaling Rules
Effects
Historical/predicted scaling
Variations (cheating)
Limits
2
Penn ESE680-002 Spring 2007 -- DeHon
Why Care?
• In this game, we must be able to predict
the future
• Rapid technology advance
• Reason about changes and trends
• re-evaluate prior solutions given
technology at time X.
3
Penn ESE680-002 Spring 2007 -- DeHon
Why Care
• Cannot compare against what
competitor does today
– but what they can do at time you can ship
• Careful not to fall off curve
– lose out to someone who can stay on
curve
4
Penn ESE680-002 Spring 2007 -- DeHon
Scaling
• Premise: features scale “uniformly”
– everything gets better in a predictable
manner
• Parameters:
 l (lambda) -- Mead and Conway (class)
 S -- Bohr
 1/k -- Dennard
5
Penn ESE680-002 Spring 2007 -- DeHon
Feature Size
l is half the minimum
feature size in a VLSI
process
[minimum feature
usually channel width]
6
Penn ESE680-002 Spring 2007 -- DeHon
Scaling
•
•
•
•
•
Channel Length (L)
Channel Width (W)
Oxide Thickness (Tox)
Doping (Na)
Voltage (V)
7
Penn ESE680-002 Spring 2007 -- DeHon
Scaling
•
•
•
•
•
Channel Length (L)
l
Channel Width (W)
l
Oxide Thickness (Tox) l
Doping (Na)
1/l
Voltage (V)
l
8
Penn ESE680-002 Spring 2007 -- DeHon
Effects?
•
•
•
•
•
•
•
•
Area
Capacitance
Resistance
Threshold (Vth)
Current (Id)
Gate Delay (tgd)
Wire Delay (twire)
Power
9
Penn ESE680-002 Spring 2007 -- DeHon
Area
 l  l/k
 A=L*W
 A  A/k2
 130nm  90nm
 50% area
 2× capacity same
area
10
Penn ESE680-002 Spring 2007 -- DeHon
Area Perspective
11
Penn ESE680-002 Spring 2007 -- DeHon
Capacity Scaling from Intel
12
Penn ESE680-002 Spring 2007 -- DeHon
ITRS 2005 Moore’s Law
13
Penn ESE680-002 Spring 2007 -- DeHon
Capacitance
• Capacitance per unit
area
– Cox= eSiO2/Tox
– Tox Tox/k
– Cox  k Cox
14
Penn ESE680-002 Spring 2007 -- DeHon
Capacitance
• Gate Capacitance
 Cgate= A*Cox
 A  A/k2
 Cox  k Cox
 Cgate  Cgate /k
15
Penn ESE680-002 Spring 2007 -- DeHon
Threshold Voltage
16
Penn ESE680-002 Spring 2007 -- DeHon
Threshold Voltage
• VTH VTH /k
17
Penn ESE680-002 Spring 2007 -- DeHon
Current
• Saturation Current
Id=(mCOX/2)(W/L)(Vgs-VTH)2
Vgs=V V /k
VTH VTH /k
W W/k
L L/k
Cox  k Cox
Id Id/k
18
Penn ESE680-002 Spring 2007 -- DeHon
Gate Delay
 tgd=Q/I=(CV)/I
 V V /k
 Id  Id/k
 C  C /k
 tgd  tgd /k
19
Penn ESE680-002 Spring 2007 -- DeHon
Overall Scaling Results, Transistor Speed and Leakage. Preliminary Data
from 2005 ITRS.
•HP = High-Performance Logic
•LOP = Low Operating Power Logic
•LSTP = Low Standby Power Logic
Leakage Current
Intrinsic Transistor
Delay, t = CV/I
10.00
(HP: standby power
dissipation issues)
(lower delay = higher speed)
1.E+00
HP
LOP
1.E-01
LSTP
1.00
CV/I (ps)
(ps)
Isd,leak (uA/um)
1.E-02
LSTP Target:
Isd,leak ~ 10 pA/um
1.E-04
0.10
HP Target:
17%/yr, historical
rate
0.01
2005
LOP
1.E-03
2007
2009
2011
2013
2015
17%/yr
rate
2017
1.E-05
Planar Bulk
MOSFETs
1.E-06
2005
2019
2007
2009
2011
Advanced
MOSFETs
2013
2015
Calendar year
Calendar Year
20
2017
2019
Resistance
• R=rL/(W*t)
• W W/k
• L, t similar
• RkR
21
Penn ESE680-002 Spring 2007 -- DeHon
Wire Delay
 twire=RC
 R -> k R
 C -> C /k
 twire -> twire
• …assuming (logical)
wire lengths remain
constant...
• Assume short wire or
buffered wire
• (unbuffered wire
ultimately scales as
length squared)
22
Penn ESE680-002 Spring 2007 -- DeHon
Power Dissipation (Static Load)
• Resistive Power
– P=V*I
– V V /k
– Id Id/k
– P P /k2
23
Penn ESE680-002 Spring 2007 -- DeHon
Power Dissipation (Dynamic)
• Capacitive
(Dis)charging
 P=(1/2)CV2f
 V V /k
 C  C /k
 P P/k3
• Increase
Frequency?
 tgd  tgd /k
 So: f  kf ?
 P  P/k2
24
Penn ESE680-002 Spring 2007 -- DeHon
…and leakage
25
Penn ESE680-002 Spring 2007 -- DeHon
[source: Borkar/Intel, Micro37, 12/04]
Intel on Leakage
26
Penn ESE680-002 Spring 2007 -- DeHon
[source: Borkar/Intel, Micro37, 12/04]
Effects?
•
•
•
•
•
•
•
•
Area
1/k2
Capacitance
1/k
Resistance
k
Threshold (Vth)
1/k
Current (Id)
1/k
Gate Delay (tgd) 1/k
Wire Delay (twire) 1
Power
1/k2 1/k3
27
Penn ESE680-002 Spring 2007 -- DeHon
ITRS Roadmap
• Semiconductor Industry rides this
scaling curve
• Try to predict where industry going
– (requirements…self fulfilling prophecy)
• http://public.itrs.net
28
Penn ESE680-002 Spring 2007 -- DeHon
MOS Transistor Scaling
(1974 to present)
S=0.7
[0.5x per 2 nodes]
Pitch
Source: 2001 ITRS - Exec. Summary, ORTC Figure
Penn ESE680-002 Spring 2007 -- DeHon
Gate
[from Andrew Kahng]
29
Half Pitch (= Pitch/2) Definition
Metal
Pitch
(Typical
DRAM)
Source: 2001 ITRS - Exec. Summary, ORTC Figure
Penn ESE680-002 Spring 2007 -- DeHon
Poly
Pitch
(Typical
MPU/ASIC)
[from Andrew Kahng]
30
Node Cycle Time:
0.7x
0.7x
Log Half-Pitch
Scaling Calculator +
1994 NTRS .7x/3yrs
Actual .7x/2yrs
Linear Time
250 -> 180 -> 130 -> 90 -> 65 -> 45 -> 32 -> 22 -> 16
0.5x
N
N+1
Node Cycle Time
(T yrs):
N+2
*CARR(T) =
[(0.5)^(1/2T yrs)] - 1
* CARR(T) = Compound Annual
Reduction Rate
(@ cycle time period, T)
Source: 2001 ITRS - Exec. Summary, ORTC Figure
Penn ESE680-002 Spring 2007 -- DeHon
CARR(3 yrs) = -10.9%
CARR(2 yrs) = -15.9%
[from Andrew Kahng]
31
ITRS 2005
32
Penn ESE680-002 Spring 2007 -- DeHon
ITRS 2003,2005 Gate/Wire Scaling
33
Penn ESE680-002 Spring 2007 -- DeHon
What happens to delays?
• If delays in gates/switching?
• If delays in interconnect?
• Logical interconnect lengths?
34
Penn ESE680-002 Spring 2007 -- DeHon
Delays?
• If delays in gates/switching?
– Delay reduce with 1/k [l]
35
Penn ESE680-002 Spring 2007 -- DeHon
Delays
• Logical capacities growing
• Wirelengths?
– No locality: Lk
– Rent’s Rule
• L  n(p-0.5)
• [p>0.5]
(slower!)
36
Penn ESE680-002 Spring 2007 -- DeHon
Compute Density
• Density = compute / (Area * Time)
 k3>compute density scaling>k
 k3: gates dominate, p<0.5
 k2: moderate p, good fraction of gate delay
– [p from Rent’s Rule again – more on Day14]
 k : large p (wires dominate area and delay)
37
Penn ESE680-002 Spring 2007 -- DeHon
Power Density
• P-> P /k2 (static, or increase frequency)
• P-> P/k3 (dynamic, same freq.)
• A -> A/k2
• P/A  P/A … or … P/kA
38
Penn ESE680-002 Spring 2007 -- DeHon
Cheating…
• Don’t like some of the implications
– High resistance wires
– Higher capacitance
– Quantum tunneling
– Need for more wiring
– Not scale speed fast enough
39
Penn ESE680-002 Spring 2007 -- DeHon
Improving Resistance
•
•
•
•
R=rL/(W*t)
W W/k
L, t similar
RkR
Don’t scale t quite as fast.
Decrease r (copper)
40
Penn ESE680-002 Spring 2007 -- DeHon
41
Penn ESE680-002 Spring 2007 -- DeHon
Capacitance and Leakage
• Capacitance per unit
area
– Cox= eSiO2/Tox
– Tox Tox/k
– Cox  k Cox
Reduce Dielectric Constant e (interconnect)
and Increase Dielectric to substitute for scaling Tox
(gate quantum tunneling)
Penn ESE680-002 Spring 2007 -- DeHon
42
Threshold Voltage
43
Penn ESE680-002 Spring 2007 -- DeHon
ITRS 2005
44
Penn ESE680-002 Spring 2007 -- DeHon
High-K dielectric Survey
Wong/IBM J. of R&D, V46N2/3P133--168
Penn ESE680-002 Spring 2007 -- DeHon
45
Intel Saturday NYT
Announcement
• Intel Says Chips Will Run Faster,
Using Less Power
– NYT 1/27/07, John Markov
– Claim: “most significant change in the
materials used to manufacture silicon chips
since Intel pioneered the modern integratedcircuit transistor more than four decades ago”
– “Intel’s advance was in part in finding a new
insulator composed of an alloy of
hafnium…will replace the use of silicon
dioxide.”
46
Penn ESE680-002 Spring 2007 -- DeHon
Wire Layers = More Wiring
47
Penn ESE680-002 Spring 2007 -- DeHon
Typical chip cross-section illustrating
hierarchical scaling methodology
Penn ESE680-002 Spring 2007 -- DeHon
[ITRS2005 Interconnect Chapter]
48
Improving Gate Delay
 tgd=Q/I=(CV)/I
 V V /k
 Id=(mCOX/2)(W/L)(Vgs-VTH)2
 Id  Id/k
 C  C /k
 tgd  tgd /k
Lower C.
Don’t scale V.
Penn ESE680-002 Spring 2007 -- DeHon
Don’t scale V:
VV
IkI
tgd  tgd /k2
49
…But
Power Dissipation (Dynamic)
• Capacitive
(Dis)charging
 P=(1/2)CV2f
 V V /k
 C  C /k
 P P/k3
• Increase
Frequency?
 f  kf ?
 P  P/k2
If not scale V, power dissipation not scale.
50
Penn ESE680-002 Spring 2007 -- DeHon
…And
Power Density
• P P (increase frequency)
• P> P/k (dynamic, same freq.)
 A  A/k2
• P/A  kP/A … or … k2P/A
• Power Density Increases
…this is where some companies have gotten into trouble…
51
Penn ESE680-002 Spring 2007 -- DeHon
Intel on Leakage
52
Penn ESE680-002 Spring 2007 -- DeHon
[source: Borkar/Intel, Micro37, 12/04]
Physical Limits
• Doping?
• Features?
53
Penn ESE680-002 Spring 2007 -- DeHon
Physical Limits
• Depended on
– bulk effects
• doping
• current (many electrons)
• mean free path in conductor
– localized to conductors
• Eventually
– single electrons, atoms
– distances close enough to allow tunneling
54
Penn ESE680-002 Spring 2007 -- DeHon
Dopants/Transistor
1000
100
1
10
100
0
500
250
130
65
Technology Node (nm)
32
16
Dopant Fluctuation
Mean Number of Dopant Atoms
10000
0.1
0.01
1000
100
Technology Nodes (nm)
55
Penn ESE680-002 Spring 2007 -- DeHon
10
Electrons
e=1.6×10-19 C
Penn ESE680-002 Spring 2007 -- DeHon
How many electrons?
56
What Is A “Red Brick” ?
• Red Brick = ITRS Technology Requirement with
no known solution
• Alternate definition: Red Brick = something
that REQUIRES billions of dollars in R&D
investment
Penn ESE680-002 Spring 2007 -- DeHon
[from Andrew Kahng]
57
The “Red Brick Wall” - 2001 ITRS vs 1999
Source: Semiconductor International - http://www.e-insite.net/semiconductor/index.asp?layout=article&articleId=CA187876
[from Andrew Kahng]
Penn ESE680-002 Spring 2007 -- DeHon
58
ITRS 2005 …
59
Penn ESE680-002 Spring 2007 -- DeHon
Conventional Scaling
• Ends in your lifetime
• …perhaps in your first few years out of
school…
• Perhaps already:
– "Basically, this is the end of scaling.”
• May 2005, Bernard Meyerson, V.P. and chief
technologist for IBM's systems and technology
group
60
Penn ESE680-002 Spring 2007 -- DeHon
Finishing Up...
61
Penn ESE680-002 Spring 2007 -- DeHon
Big Ideas
[MSB Ideas]
• Moderately predictable VLSI Scaling
– unprecedented capacities/capability growth
for engineered systems
– change
– be prepared to exploit
– account for in comparing across time
– …but not for much longer
62
Penn ESE680-002 Spring 2007 -- DeHon
Big Ideas
[MSB-1 Ideas]
• Uniform scaling reasonably accurate for
past couple of decades
• Area increase k2
– Real capacity maybe a little less?
• Gate delay decreases (1/k)
• Wire delay not decrease, maybe increase
• Overall delay decrease less than (1/k)
63
Penn ESE680-002 Spring 2007 -- DeHon