Академический Документы
Профессиональный Документы
Культура Документы
Outline
Introduction Transistor performance
How does transistor performance change as technology is scaled down to nanoscale* nanoscale dimensions
*The
1995
2000
2005
2010
2015
Vdd [V]
0.1 1970 1975 1980 1985 1990 Year 1995 2000 2005 2010 2015
Transistor performance
Transistor performance metrics for analog and RF design:
Transconductance Intrinsic gain (gm/gds) Capacitances Maximum operating p g frequency q y Efficiency (gm/Id) Linearity Noise Mismatch Gate leakage
Transconductance is, for f all practical purposes, independent of the technology node F a velocity For l it saturated t t d channel: h l
gm [S]
1.E-06
I d = WcoxVeff vs g m = Wcox vs
Minimum gate length, same W /L V ds = 1/2V DD
1.E-07
1.E-08
W L; cox
65 nm 90 nm 0.13 um 0.18 um
Intrinsic gain is reduced as technologies are scaled down down. At the 65nm node the maximum intrinsic gain is reduced by more than 80% compared to the g gain at the 0.18m node.
gm /gds
1.E-09
1.E-08
1.E-07
1.E-06
1.E-05
1.E-04
1.E-03
Transistor capacitances
1.1 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 90nm 0.13um 0.18um
Cgs Cdb
Minimum gate length, same W /L ; V ds = 0
Capacitances extracted from NMOS having minimum drain and source areas for the given W/L (~3).
Capacitance e [fF]
gm 2(C gs + C gd + Cdb )
2.5E-04
(1)
65 nm
1.0E+11
90 nm 0.13 um
1.0E+10 1.0E+09
0.18 um
1 5E-04 1.5E 04
Id [A]
1.0E-04
5.0E-05
0.0E+00 1.E+06
1.E+07
1.E+08
1.E+09
1.E+10
1.E+11
f T [Hz]
65 nm 5.0
90 nm
0 13 um 0.13
0 18 um 0.18
4.0
3.0
20 2.0
By scaling down a potential improvement in speed can be obtained compared to realization in a 0.18m technology
1.E-03
1.0
Minimum gate length, same W /L ; V ds = 1/2V DD Speed potential normalized to the 0.18m node
0.0 0 0 1.E-10 1.E-09 1.E-08 1.E-07 1.E-06 1.E-05 1.E-04
gm /ID [ [1/V]
20 15 10 5 0 1.0E-12
1.0E-10
1.0E-08
1.0E-06
1.0E-04
1.0E-02
The transistor is most efficient (maximum bang for the buck) in weak inversion Maximum value is nearly independent of technology
9 Trond Ytterdal, Department of Electronics and Telecommunication
Linearity (1)
The Taylor expansion of the small signal drain current of a MOS transistor can be written as: 2 id = g m v gs + a 2 g m v gs + a3 g m v 3 gs + L
2 3 + g ds vds + a 2 g ds vds + a3 g ds vds +L 2 3 + g mb b vbs b + a 2 g mb vbs b + a3 g mb vbs b +L 2 2 + a 2 g m g ds v gs vds + a32 g m g ds v gs vds + a3 g m 2 g ds v gs vds +L 2 2 + a 2 g m g mb v gs vbs + a32 g m g mb v gs +L vbs + a3 g m 2 g mb v gs vbs 2 2 + a 2 g ds g mb vds vbs + a32 g ds g mb vds vbs + a3 g ds 2 g mb vds vbs +L
+ a3 g m g ds g mb v gs vds vbs + L
Here, the coefficients are the higher order derivatives of the total drain current with respect to one or more of the control voltages. voltages For example:
1 2Id a2 g m = 2 2 Vgs
10 Trond Ytterdal, Department of Electronics and Telecommunication
Linearity (2)
1.0E-03 1.0E-04 1.0E-05
|a 2gds| [S/V V]
Second order nonlinearity in channel conductance (gds) increases as t h l technology i is scaled down.
1.E-08 1.E-07 1.E-06 1.E-05 1.E-04 1.E-03
1.E-09
Id [A]
1.0E-02
1 0E 03 1.0E-03
a 2g gm [S/V]
1.0E-04
1 0E 05 1.0E-05
1.0E-06
1.0E-07 1 0E 07 1.0E-10
1.0E-09
Id [A]
11 Trond Ytterdal, Department of Electronics and Telecommunication
Noise
Equivalent noise PSD (power spectral density) referred to the gate of a transistor connected in a common-source configuration*:
K 2 = 4 k BT + Vng g m WLcox f ; K process dependent
Si-SiO2 interface noise Channel noise Increases at low current Minimum in weak inversion for given current Traditionally [1], the following expression is used for :
2n / 3 in strong inversion = n / 2 in weak inversion
*Transform
12
60
40
20
13
WLf
10log(Id 2/(Ind 2Agate )) [dB]
80
65nm
gm 4k BTK I d
75
70
65
60
55
14
Mismatch (1)
Mismatch of two matched transistors identical by design can be characterized by:
VT = VT 1 VT 2 with VT = (VT ) = AVT / WL / = 2(1 2 ) /( (1 + 2 ) with = ( /) = A / WL
AVT and A depends on process, layout and more Based on the quadratic MOS model, one can derive the following well known formulas:
Same gate voltage:
I d I d gm 2 = + VT Id
2
15
Mismatch (2)
Mismatch per unit area decreases as technologies are scaled down
From theory and ideal scaling:
Gate area required for any given (VT) normalized to the Lmin = 70nm
1000
100
10
10
1 0.01 0.1 1 10
1 0.01
0.1
10
VG
IG
0.5
1.5
g-flavor technologies
17 Trond Ytterdal, Department of Electronics and Telecommunication
vg ig
ic
itunnel gtunnel
Cox
At a given frequency the impedance of the two branches have the same magnitude g and the currents ic and itunnel will be equal q in magnitude. This frequency is given by
gateCox = gtunnel 1 gtunnel f gate = 2 Cox
Above this frequency, frequency the total gate current is dominated by ic as in the traditional case.
18 Trond Ytterdal, Department of Electronics and Telecommunication
2 ox f gate t K fg f Vgs e
t (V gs 13.6)
Here, tox is the oxide thickness in nm and Kfg is 1 1.5 51016 for NMOS and 0.51016 for PMOS. A plot of this expression is shown to the right for three different technologies.
10-5 10-6 10-7 10-8 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Veff [V]
19
Accuracy
20
FOM definition
One commonly used Figure-of-merit (FOM) definition:
FOM = P DR 2 f
Here, P is the power consumption, DR is the dynamic range and f is the highest signal frequency that can be processed by the circuit Small is GOOD.
21
where, Vpp is the maximum peak-to-peak value of the signal, Vn_rms is the RMS noise voltage, v is the voltage l efficiency ffi i defined d fi d by b Vpp/VDD.
*The
22
where is a yield parameter and (Vos) is the standard deviation of a critical offset voltage in the circuit. We observe for both ( (2) ) and (3) ( ) that DR scales with VDD and v For the rest of the talk we assume that DR is limited by noise on the lower side
23 Trond Ytterdal, Department of Electronics and Telecommunication
(vVDD ) 2
2 8Vn _ rms
(4)
24
Vpp
vout
vin VSS
vout 1/f
1 1 1 1 2 fq = VDD fV pp C = fV pp C i i v i Here, i is the current efficiency of the transconductor and q/i is the charge transferred from the power supply in one period and. and We have assumed that VSS = 0
SNR:
SNR =
k T 2 V pp = 8 B SNR k BT / C C
2 V pp /8
Combining the two equations we get the power required per pole versus SNR:
P=
Based on [1] and [4]
25
8k BT f SNR i v
(6)
Theoretical FOM
Rewriting (6) in terms of FOM:
P 8k BT P 4 k BT = = 2 f SNR i v i v f DR 4k T FOM = B i v
Hence, H one i important question i becomes: b How does i and v depend on technology?
26 Trond Ytterdal, Department of Electronics and Telecommunication
27
2 EN O B Fs FM = Pdiss
Philips ISSCC04
1.0E+13
State-of-the-art 2004
1997 1998 1999 2000 2001 2002 2003 2004 Selection from Walden
Gaggl, ISSCC04
This design(03)
Bjrnsen ESSCIRC05[5]
Kelly, ISSCC01 nAD12110-18(03) AD9245(03) H Hernes ISSCC04 nAD10120-13(03)
1.0E+12
1.0E+11
Siragusa ISSCC04
1.0E+10
1 MHz
10 MHz
100 MHz
1 GHz
From [5]
28 Trond Ytterdal, Department of Electronics and Telecommunication
65nm
0.25
90nm 0.13um
where Vdsat is the saturation voltage of our transistors To keep v unchanged when we scale down technology, Vdsat must be scaled down at the same rate as VDD.
0.20
Vdsatt [V]
0.15
0.10
0.05
0.00 -0.6
-0.4
-0.2
0.2
0.4
0.6
Veff [V]
(8)
Here gm is the transconductance of the input transistor and C is the load capacitance. Writing the square of the noise RMS voltages as: k BT 2 , Vn _ rms = C Inserting this equation in (2), (2) solving for C and inserting in (8) yields
GBW = g m (vVDD ) 2 8k BT DR
2
(9)
Hence, if DR and gm is kept constant, the speed of the circuit decreases with the square of the supply voltage.
30
Power consumption doubles i decreases FOM degrades Downscaling will not help
31
1.0E-10
1.0E-08
1.0E-06
1.0E-04
1.0E-02
1.5E-04
1.0E-04
5.0E-05
0.0E+00 1.E+06
1.E+07
1.E+08
1.E+09
1.E+10
1.E+11
f T [Hz]
fTmax@90nm
65 nm 90 nm 0.13 um 0.18 um
1.E-09
1.E-08
1.E-07
1.E-06
1.E-05
1.E-04
1.E-03
33
Maximizing FOM
Some guidelines for maximizing FOM in nanoscale CMOS circuits: Circuit level:
Maximize v by using as low Veff ff as possible Identify fTmax in your technology. Stay well below this frequency to maximize i * Use U as hi high h supply l voltage lt as possible ibl (maybe ( b even higher than allowed by the foundry)
Architecture level:
Avoid high gain requirements (use, for example, gain calibration) ) Go to interleaved architectures if speed requirements cause you to move dangerously close to fTmax*
*or
34
Summary
Back to near constant voltage scaling (at least for a while) Back to near constant voltage scaling
Reliability issues: Have to verify circuits versus age
Speed of transistors SNR/DR Mismatch Linearity Potential for FOM improvement in some circuits by scaling down technology*
*One
example is ADCs where transistors are biased in regions of low current efficiency
35 Trond Ytterdal, Department of Electronics and Telecommunication
List of symbols
Parameter Agate Cgs Cgd Cdb cox Cox f gm gds Id kB L Lmin Gate area coxW/L Gate-source capacitance Gate-drain capacitance Drain-bulk capacitance Oxide capacitance per gate area O id capacitance Oxide it Frequency Transconductance Channel conductance Drain current Boltzmanns l constant Gate length Minimum g gate length g Description Parameter n T tox VDD Vds Veff Vgs vs VT W Mobility Subthreshold ideality factor Absolute temperature Oxide thickness Supply voltage Drain-source voltage Vgs VT Gate-source voltage Saturation velocity Threshold voltage Gate width Description
37
References
[1] E.A. Vittoz, Low Power Design: Ways to Approach the Limits, SolidState Circuits Conference, 1994. Digest of Technical Papers. 41st ISSCC, pp 14 pp. 14-18, 18 Feb Feb. 1994. 1994 [2] K. Bult, Analog Design in Deep Sub-Micron CMOS, in Proc. ESSCIRC 2000, pp. 11-17. [3] A. Annema et al., Analog Circuits in Ultra-Deep-Submicron CMOS, IEEE J. of Solid-State Circuits, vol. 40, no. 1, Jan. 2005. [4] E.A. E A Vittoz, Vittoz Low Low-power power Low Low-Voltage Voltage Limitations and Prospects in Analog Design, in Analog Circuit Design, Editors R.S. van de Plaasche, W.M.C. Sansen, and J.H. Huijsing, Kluwer Academic Publishers, 1994. [5] J. Bjrnsen, . Moldsvor, T. Sther, T. Ytterdal, A 220mW 14b 40MSPS Gain Calibrated Pipelined ADC, in Proc. European Solid-State Circuits Conference f (ESSCIRC) ( ) 2005, Grenoble, Sept. p 12-14, pp pp. 165-168, 2005.
38