UNIVERSITY OF CALIFORNIA College of Engineering Electrical...
Transcript of UNIVERSITY OF CALIFORNIA College of Engineering Electrical...
UNIVERSITY OF CALIFORNIA College of Engineering
Department of Electrical Engineering and Computer Sciences
Jan M. Rabaey Homework 1 EECS 247 Solution Spring 2005
1. SPICE models
Vthp_extrap2 0.713V=Vthn_extrap2 0.701V=
Vthp_extrap2 Vgsp 0A( ):=Vthn_extrap2 Vgsn 0A( ):=
Vgsp Idp( ) 1
mp2
Idp yp2−( )⋅ xp2+:=Vgsn Idn( ) 1
mn2
Idn yn2−( )⋅ xn2+:=
yp2 11.46µA:=xp2 0.958V:=mp2 46.69 106−S⋅:=yn2 34.87µA:=xn2 0.882V:=mn2 192.332 10
6−S⋅:=
For |Vbs|=0.5V
Vthp_extrap1 0.837V=Vthn_extrap1 0.795V=
Vthp_extrap1 Vgsp 0A( ):=
Body-effect parameter, γγγγ
Vth Vth0 γ 2− φF VSB+ 2− φF⋅−( )+=
φFp 0.3V:= VSB 1V:=
For |Vbs|=1V
mn1 180.42 106−S⋅:= xn1 0.943V:= yn1 26.66µA:= mp1 44.19 10
6−S⋅:= xp1 1.059V:= yp1 9.795µA:=
Vgsn Idn( ) 1
mn1
Idn yn1−( )⋅ xn1+:=Vgsp Idp( ) 1
mp1
Idp yp1−( )⋅ xp1+:=
Vthn_extrap1 Vgsn 0A( ):=
S 38mV/decade=S 24mV/decade=For 77K:
S 97mV/decade=S 93mV/decade=For 298K:
PMOS NMOS
Slope factor
b) Subthreshold slopes
Values reported in the model file is 0.55 and 0.63 for NMOS and PMOS respectively
γp 0.615 V=γn 0.483 V=
γp
Vthp_extrap1 Vthp0_extrap−
2− φFp⋅ VSB+ 2− φFp⋅−( ):=
γn
Vthn_extrap1 Vthn0_extrap−
2− φFn⋅ VSB+ 2− φFn⋅−( ):=
VSB 1− V:=
φFp 0.443V:=andφFn 0.443V:=
Solving the above threshold voltage equation for the two extrapolated threshold voltage with source bulk reverse bias
of 1V and 0.5 V respectively, we get
Vth Vth0 γ 2− φF VSB+ 2− φF⋅−( )+=
Zoomed-in plots of the log(Id) at low Vgs:
Leakage Current at 0Vgs:
NMOS PMOS
For 298K: Ist 5.67pA= Ist 3.156pA=
For 77K: Ist 45.9zA= Ist 112.15zA=
||||
ds
th
VV
DIBL∆∆=
From the above plots the DIBL factors are calculated to be 5e-6 and 2e-3 for NMOS and PMOS respectively.
2. Technology scaling
Eq. (1)tp_035_FO1 tp0_035 1f035_FO1
γ+
⋅=
f035_FO1
CL_035_FO1
Cin_035_FO1
:=CL_035_FO1 4.4362f F⋅:=Cin_035_FO1 2.577f F⋅:=
tp_035_FO1 64.28p s⋅=tp_035_FO1
Tringosc_035
10:=Tringosc_035 642.798p s⋅:=
Intrinsic delay, tp
Cin_035 3.33f F⋅:=Total input capacitance,
signal OP("/M2" "??")
betaeff 369.7u
cbb 250.1a
cbd -397.9z
cbg -247.1a
cbs -2.609a
cdb -3.826z
cdd 173.3a
cdg -173.3a
cds 16.43z
cgb -250.1a
cgd -172.9a
cgg 593.8a
cgs -170.8a
cjd 732.5a
cjs 0
csb -5.739z
csd -32.08y
csg -173.3a
css 173.3a
signal OP("/M3" "??")
betaeff 169.8u
cbb 439.7a
cbd -819.5a
cbg -901.1z
cbs 380.7a
cdb -174.3a
cdd 1.565f
cdg -1.369f
cds -21.96a
cgb -91.02a
cgd -1.31f
cgg 2.739f
cgs -1.338f
cjd 2.836f
cjs 0
csb -174.3a
csd 564.2a
csg -1.369f
css 979.2a
a) For 0.35µm inverter
2
Esw_035 38.11f J⋅:=Active energy consumption
Simulated Esw
tp0_035 28p s⋅:=γ 1.3354:=
Solving Eq. 1 and 2 for tp0_035 and γ
Eq. (2) tp_035_FO4 tp0_035 1f035_FO4
γ+
⋅=tp_035_FO4 =
f035_FO4
CL_035_FO4
Cin_035_FO4
:=CL_035_FO4 12f F⋅:=Cin_035_FO4 3.33f F⋅:=
tp_035_FO4 0.5 tpHL_035_FO4 tpLH_035_FO4+( )⋅:=tpLH_035_FO4 125.97p s⋅:=tpHL_035_FO4 81.63p s⋅:=
Esw_025_est 15.623 f J⋅=Esw_025_est
Esw_035
S U2
⋅
:=
Intrinsic energy scales as 1/SU2
tp0_025_est 20 p s⋅=tp0_025_est
tp0_035
S:=
Intrinsic delay scales as 1/S
Cin_025_est 2.379 f F⋅=Cin_025_est
Cin_035
S:=
Input capacitance scales as 1/S,
U 1.32=U3.3V
2.5V:=
Voltage scaling parameter,
S 1.4=S0.35µ m⋅
0.25µ m⋅:=
Technology scaling parameter,
With generalized scaling,
Predicting 0.25um values
Simulated parameters for 0.25µm inverter:
Cin
signal OP("M3" "??")
betaeff 152.5u
cbb 427.2a
cbd -654.7a
cbg -4.306a
cbs 231.8a
cdb -175.8a
cdd 1.293f
cdg -1.128f
cds 10.85a
cgb -75.49a
cgd -1.102f
cgg 2.261f
cgs -1.083f
cjd 2.531f
cjs 0
csb -175.8a
csd 463.8a
csg -1.128f
css 840.3a
signal OP("M2" "??")
betaeff 428.8u
cbb 210.9a
cbd -946.7z
cbg -208.7a
cbs -1.243a
cdb -106.9y
cdd 205.4a
cdg -205.4a
cds 623.7y
cgb -210.9a
cgd -204.4a
cgg 619.5a
cgs -204.1a
cjd 879.9a
cjs 0
csb -160.4y
csd -3.865y
csg -205.4a
css 205.4a
For 0.25µm inverter, input capacitance,
Cin_025 2.88f F⋅:=
Intrinsic delay, tp0
Tringosc_025 582.328p s⋅:= tp_025_FO1
Tringosc_025
10:=
Cin_025_FO1 2.431f F⋅:= CL_025_FO1 3.957f F⋅:= f025_FO1
CL_025_FO1
Cin_025_FO1
:=
tp_025_FO1 tp0_025 1f025_FO1
γ+
⋅= Eq. (3)
tpHL_025_FO4 67.59p s⋅:= tpLH_025_FO4 106.286p s⋅:= tp_025_FO4 0.5 tpHL_025_FO4 tpLH_025_FO4+( )⋅:=
Cin_025_FO4 2.88f F⋅:= CL_025_FO4 11.22f F⋅:= f025_FO4
CL_025_FO4
Cin_025_FO4
:=
tp_025_FO4 tp0_025 1f025_FO4
γ+
⋅= Eq. (4)
Solving Eq. 3 and 4 for tp0_025
γ 1:= tp0_025 22p s⋅:=
Active energy consumption, Esw
Active energy consumption
Esw_025 24f J⋅:=
Comparison of results:
Extracted 0.35um parameters Predicted 0.25um parameters Extracted 0.25um parameters
Cin_035 3.33 f F⋅= Cin_025_est 2.379 f F⋅= Cin_025 2.88 f F⋅=
tp0_035 28 p s⋅= tp0_025_est 20 p s⋅= tp0_025 22 p s⋅=
Esw_035 38.11 f J⋅= Esw_025_est 15.623 f J⋅= Esw_025 24 f J⋅=
Comments:
The predicted parameters matches the extracted parameters to some degree but not fully. This is because not every
thing scale as a factor of S/U. For instance, Vth doesn't scale fully as 1/U as supply voltage does and the junction
depth also does not scale with decreasing channel length. Parasitic resistances are also ignored in the scaling
theory. These caveats naturally contribute to the discrepancies we see, especially in the switching energy.
3. Design scaling
a) For the PDA SOC LP driver scaling predicted by ITRS, the parameters of technology node, supply voltage, clock frequency and average active power normalized to the 2003 references are plotted in the following graph.
ITRS LP SOC Design Scaling
0
1
2
3
4
5
6
2000 2005 2010 2015 2020
Year
Normalized Parameters
Tech node
Vdd
Clock freq
Active power
b) Main requirement of this problem is on understandings of active power calculation in CMOS circuits. At the same time the exploration on modeling the impacts of different circuit design strategies on power consumption is encouraged.
Active power consumption of a design is a function of activity factor (α), supply voltage (Vdd), clock frequency (É) and the charge capacitances (CL). CL is approximately inversely proportional to the technology scaling (due to both tox and gate dimension scalings), and proportional to the gate count in a design.
A basic model of calculating the active power is
SAreaS
AreaS
C diedie
L ⋅=∝ 2/11
, SAreafVfVCP dieddddLactive ⋅⋅∝∝ 22 αα
Here S is the technology scaling factor. It is estimated by ITRS that die size increases on average by 10% per node. As the activity factor varies among different designs and applications, in this model we assume the same level of α. With the ITRS scaling data in a), a basic model of the active power is as following:
Source 2003 2006 2009 2012 2015 2018 Tech node ITRS 101 90 65 45 32 22 Vdd ITRS 1.2 1 0.8 0.6 0.5 0.4 Clock freq ITRS 300 450 600 900 1200 1500 Die size ITRS 1 1.1 1.21 1.331 1.4641 1.6105 Modeled active P Vdd^2*f*die/tech 4.277 5.5 7.148 9.583 13.73 17.57
As a result of the ever-increasing integration, the dynamic power grows about 30% at each technology generation. Effective power control technique is required to minimize the power
gap between technology scaling and SOC low power design requirement, which is also reported in the ITRS analysis. Such power control techniques include parallel architecture, higher memory to logic ratio, CMOS manufacture process engineering and even migration to non-classical CMOS structures. The goal is to keep the SOC LP design power at a constant level.
An example of a quantitative model considering the impact of applying parallel architecture on chip active power is as following:
Source 2003 2006 2009 2012 2015 2018Transistor count (mil) 41*die/(tech/101)^2 41 56.7982 119.8 274.9 597.99 1391.7SRAM T count 2X per node 16 32 64 128 256 640Core logic T count 2~4X per node 2 8 32 64 256 512Periph logic T count total T - SRAM T - core logic T 23 16.7982 23.78 82.903 85.995 239.7Active SRAM T SRAM T * α(SRAM) 0.064 0.09051 0.128 0.181 0.256 0.4048Parallelism Assumed 1 1 2 2 2 3Equivlant active logic T α(logic)*(Core T / Para^2 + Periph T) 2.5 2.47982 3.178 9.8903 14.999 29.659Modeled active P Vdd^2*f*ActiveT(SRAM+logic)*tech 111.87 104.098 82.52 146.84 146.45 158.73
This model uses the ITRS prediction data of chip memory size, core and peripheral logic size scaling, as well as the SRAM activity factor scaling. By assuming 2X parallelism after 65nm node and 3X after 22nm node, we can approximately keep the active power constant. However notice that with a supply voltage of 0.4V at 22nm node, the 3X parallelism is not very realistic. At the same time a memory size (640 million transistors) larger than the scaled value at 22nm node was used to help the power suppression. All these indicate the big power control challenge in future CMOS design scaling.
4. CMOS scaling
a) When scaled below the 90nm node, traditional CMOS structure faces the problems of high leakage current, pronounced short channel effect, large process variation, and reduced source/drain conductivity under low supply voltage. To find a device structure that has improved performance and sustainable scalability is the motivation for exploring the non-classical CMOS structures.
b) The non-classical CMOS structures are reviewed as following.
Transport-Enhanced MOSFET: A strained semiconductor is manufactured below the gate. Advantage: the strained silicon increases the mobility of the carriers in the channel,
resulting in an increase of the saturation current. Disadvantage: High manufacturing cost (requires epi-deposition and SOI)
UTB SOI MOSFET: A thin transistor body is placed over oxide. Advantage: Good isolation and electrostatics. Good control of the channel by the gate in
the off state. Disadvantage: High manufacturing cost (requires SOI). Vth adjustment difficult. Source/Drain Engineered MOSFET 1) Source and drain are metallic from silicidation and are isolated from other transistors
by a Shocttky diode. Advantage: Low source / drain resistance. No longer need shallow junction S/D.
Disadvantages: The Shottky barrier is not as high as diode barrier. It is costly to manufacture (need SOI). The Source and Drain must be silicided.
2) A second type has a reduced Gate to Source and Gate to Drain overlap to reduce parasitic capacitance.
Advantage: Low gate to drain capacitance. Disadvantage: The channel resistance would increase since not all the channels are
inverted. Multi-Gate MOSFET Control the channel by either multiple gates or a wrap-around gate. 1) Wrap-around Gates: Advantage: Higher drive current due to dynamic Vth. Good control of the channel in the
off state. Disadvantage: Less control over channel size. 2) Double Gate MOSFETS Advantages: High drive current. Very adjustable Vth. Allows for new logic styles and
new logic gate. Disadvantages: High manufacture cost. Transistor width can not be continuously tuned
in design.