Professional Documents
Culture Documents
Abstract— The present paper proposes the simple de- pads and a device under test (DUT) degrade accuracy of
embedding method for 110-GHz on-wafer S-parameter mea- de-embedding. The issue will be described in Section III-
surements. While conventional de-embedding methods re- C. Purpose of the present paper is development of more
quire two or more dummy patterns, our method only uses
a through pattern and can perform accurate de-embedding versatile method than previous works: wider range of
up to 110 GHz. Differences among the proposed method and applications, wider frequency-range and smaller area.
previous through-only de-embedding methods are that our So-called open-short and short-open de-embedding
method is available for various device under tests and higher methods and these limitations are reviewed in Section
frequencies. II for clarifying difficulties of mm-wave de-embedding.
Index Terms— de-embedding, mm-wave band, on-wafer
measurement, S-parameter measurement, RF CMOS Our through-only de-embedding method is described in
Section III. Accuracy of de-embedding depends on model-
I. I NTRODUCTION accuracy of dummy patterns such as an open, a short and
Millimeter-wave CMOS circuits have been becoming a through. As a result of evaluations of a through model
a hot research topic today [1], [2]. Big challenges for up to 110 GHz and its characteristics, the present work
designers are obviously measurements and modeling of applies the procedure in [12] and presents a 110-GHz
circuit elements in mm-wave bands such as 110 GHz. de-embedding method for various on-chip devices. It will
On-wafer S-parameter measurements are necessary for be shown that our method can perform very accurate de-
characterizing on-chip devices and circuits, and one of the embedding up to 110GHz.
reasons for difficulty of high-frequency on-wafer measure-
ments is parasitics of probing pads. A common way to II. C ONVENTIONAL M ETHODS
remove parasitics from measured S-parameter is pad de- Limitations of conventional short-open and open-short
embedding. Commonly-used de-embedding methods em- de-embedding methods are discussed. Figure 1 shows
ploy open and short on-chip standards (dummy patterns). micrographs of short and open patterns and these con-
De-embedding procedures have become increasingly more nections. Test patterns are fabricated by 0.18 µm CMOS
complex for improving de-embedding accuracy [3], [4], process with five metal layers. A metal is aluminum, and
however this method has a limitation in an applicable a dielectric is silicon dioxide. Ground-signal-ground-type
frequency [5]. The other issue is that many kinds of and a pads are used, and pad pitches are 150 µm. A short-open-
number of dummy patterns are required in de-embedding load-through calibration using a commercial impedance
methods using open and short patterns [3]–[5]. Area costs standard substrate is performed before measurements, and
of test-CMOS-chips for evaluating device characteristics calibration reference planes are at edges of probe tips.
in mm-wave bands are extremely high because these are The short-open de-embedding starts with the short cor-
fabricated by using sub-100-nm CMOS processes such as a
65 nm process. A large number of de-embedding dummy- 150µm
150µm
G G G G
patterns mean huge test- and design-costs.
We propose the through-only de-embedding method up S S S S
to 110 GHz for on-chip passive devices such as transmis-
sion lines, inductors, RF capacitors, resistors and so on. G G G G
Only a through dummy-pattern is required to perform de- S S
embedding so that area of dummy patterns is saved. Sev- pad pad pad pad
eral through-based de-embedding methods have been dis- G G
cussed from measurements of transmission lines [6]–[10]. Zshort ( Yshort ) Zopen ( Yopen )
A paper [11] has examined through-only de-embedding (a) Short pattern. (b) Open pattern.
for on-chip devices other than transmission lines, however Fig. 1. Micrographs and connections of short and open patterns.
series parasitic resistance and inductance among probing
Short-open method
be applied to high-frequency de-embedding. We propose
1
the de-embedding method using the through. Signals are
0 Open-short method
terminated in the opposite terminal (port), and signal prop-
agations in the through pattern seem to be controllable.
-1
Measured data III. T HE T HROUGH -O NLY D E -E MBEDDING M ETHOD
-2 UP TO 110 GH Z
A. An Equivalent Circuit Model of the Through
-3
0 20 40 60 80 100 Accuracy of de-embedding is almost determined by
Frequency [GHz] equivalent circuit models of dummy patterns. De-
Fig. 3. Short-open and open-short de-embedding results. The through
embedding makes sense only if equivalent circuit models
pattern is used for checking self-consistency. of test and dummy patterns exactly describe real character-
istics of them. Reliability of de-embedded data is degraded
rection as follows: by undesirable parasitic components that are not consid-
Yso = (Zmeas − Zshort )−1 − (Zopen − Zshort )−1 . ered in de-embedding procedures as described in Sec. II.
First, the equivalent circuit model of the through should
The procedure of the open-short de-embedding is be discussed for performing accurate de-embedding.
Yos = ((Ymeas − Yopen )−1 − (Yshort − Yopen )−1 )−1 , Figure 5 is the equivalent circuit of the through shown
in Fig. 2, and Fig. 6 shows S-parameters of the model and
where Yso and Yos are Y-matrixes of the DUT de- measurements. The simple through model and measured
embedded by short-open and open-short methods, respec- S-parameter are well matched as shown in Fig. 6. The
tively. Zmeas and Ymeas are Z-matrix and Y-matrix of test through can be modeled as a π-type lumped constant
circuits with pads and the DUT. circuit because a length of a through line is less than
A through pattern shown in Fig. 2 is used to check 100 µm which is much shorter than the wave length at
adequacy of de-embedding. The through pattern consists 110 GHz. As a length of the through line is very short,
of two pads, and these pads are directly connected each signal pads may be capacitively-coupled each other. Signal
other as shown Fig. 2. Self-consistency of de-embedding pads should be separated as much as possible under the
can be evaluated by removing parasitics from the through condition that the through line can be treated as a lumped
pattern as Zmeas = Zthru (Ymeas = Ythru ) in above constant circuit.
equations. An ideal through, i.e. no reflections and no
attenuation, remains if a de-embedding method is self- B. The De-Embedding Procedure
consistent. Discrepancy between the ideal through and the The present work found that characteristics of the
de-embedded through means a de-embedding error. through can be described as a π-type equivalent circuit
Figure 3 shows de-embedded S-parameters and mea- up to 110 GHz. We propose to apply the procedure in
sured data of the through. Magnitude of de-embedded S21 [12] to 110-GHz de-embedding for various DUTs. The
is around 0 dB below 40 GHz. However, at over 40 GHz, procedure of the through-only de-embedding method is as
0 0
Magnitude [dB]
Magnitude [dB]
S 21 0
Phase [deg.]
-10 Model -1 S 21
Measurement 60
-20 -2 S11 0.15
-30 S11 -3 120 0.10
Magnitude [dB]
S 21
-40 -4 180 0.05
0.01 0.1 1 10 100 0.01 0.1 1 10 100
Frequency [GHz] Frequency [GHz] 0
Fig. 6. S-parameters of the through pattern. Parameters are based on -0.05
measurements and a model in Fig. 5. -0.10 S 12
Tmeas S 21
Phase [deg.]
(a) A test pattern with pads and a device under test (DUT).
0
PAD left PAD right Z 2 /2 Z 2 /2
left right S 12
Tpad Tpad Z1 Z3
-1.0
0 20 40 60 80 100
Tthru (Ythru ) Frequency [GHz]
(b) A through and a circuit model. Fig. 8. De-embedded results of the through pattern. The proposed
Fig. 7. Connections and a circuit model. through-only de-embedding method is applied. The maximum magnitude
of S11 is -33.7 dB.
Magnitude of S 11 [dB]
Measured data Short-open method
Mangan s
15 method ’
S S (w/o de-embedding)
-10 Mangan’ s
CL CL
10 method
5fF G G -20
5 C L=1fF DUT
0 -30
Proposed method
-5
0.1 1 10 100 -40 Our method
R L [Ω]
Open-short method
Fig. 9. Simulations of de-embedding errors. Signal frequency is -50
100 GHz. 0 20 40 60 80 100
Frequency [GHz]
RL is the known value, thus de-embedding errors can be
estimated from difference between RL and extracted resis- 0
Our method
Magnitude of S 21 [dB]
′ Mangan’ s method Open-short
tance RL by de-embedding. Figure 9 shows de-embedding method
-2
errors. The Mangan’s method has large errors when series
impedance among the DUT and pads becomes dominant -4
as compared to impedance of the DUT. In the simulation, Measured data
our method has no errors as shown in Fig. 9 and can be -6 Short-open
applied to devices other than transmission lines. method
Accuracy of our method is discussed from measure- -8
ments. Figure 10 shows de-embedded S-parameters of -10
an on-chip transmission line. A co-planar structure is 0 20 40 60 80 100
measured. A line width, a ground width and a space among Frequency [GHz]
a signal line and the ground are 2.0 µm, 10.0 µm and (a) S-parameters.
2.4 µm, respectively. Designed (loss less) characteristic
200
impedance is 50 Ω. A line length is 1 mm. M5 layer
(top layer) is used. S-parameters de-embedded by our
150
method and the Mangan’s method [9] are almost the same. Short-open method
|Z 0 | [Ω]