Multigigabit Bottleneck
Multigigabit Bottleneck
Multigigabit Bottleneck
I. I NTRODUCTION
The ubiquitous adoption of wireless local area networks
(WLANs) and cellular networks over the last decade has been
propelled by the economies of scale associated with digital
receiver architectures that leverage Moores law for lowpower, low-cost implementations. An integral component of
such receiver architectures is the Analog-to-Digital Converter
(ADC), which converts the received signal into digital format,
typically with a precision of 8-12 bits. Operations such as
synchronization, equalization, demodulation and decoding are
then performed in the digital domain, greatly enhancing the
flexibility available to the designer. We would like to scale this
mostly digital paradigm to multi-Gigabit speeds, in order
to enable mass market multi-Gigabit WLANs and wireless
personal area networks (WPANs) based on large amounts of
unlicensed bandwidth available for Ultra-wideband (UWB)
communication (3.1-10.6 GHz in the US) and millimeter wave
communication (57-64 GHz in the US). The bottleneck in such
scaling becomes the ADC: high-speed, high-precision ADC
is either unavailable, or is too costly and power-hungry [1].
In this paper, we discuss two approaches for circumventing
this bottleneck by relaxing the requirements on the ADC.
1 This work was supported by the National Science Foundation under grants
ECS-0636621 and CCF-0729222. We wish to acknowledge collaborations
with our colleagues Dr. Munkyo Seo, Prof. Onkar Dabeer, and Prof. Mark
Rodwell.
22
Unquantized
2bit quantization : optimal
2bit quantization : 4 PAM / ML
2.5
1.5
0.5
0
4
10
12
14
16
SNR (dB)
Fig. 2.
Capacity with 2-bit quantization: 4-PAM with ML quantization
provides close to optimal performance.
23
I
Downconvert
with fixed
Q
Received
Local Oscillator
1
0
4
6
1-bit ADC
Branch 2
(a)
Receiver Architecture
(b)
8-Sector Phase
Quantization
Offset
Correction
(a)
Offset
Estimation
DSP
Low-Precision
ADC
1-bit ADC
1-bit ADC
Consider linear modulation over a LoS link with ideal timing synchronization. The receiver downconverts the received
passband signal with a fixed local oscillator which has an offset f relative to the frequency of the incoming carrier. Prior
to quantization, the Nyquist-rate complex baseband samples
are given by
Offset
Estimation
1-bit ADC
Passband
Signal
High-Precision
ADC
Branch 1
Demodulator
Offset
Correction
(b)
where {Xi [n], n = 1, ..., L} are the symbols in the ith block,
{Ni [n]} is complex AWGN, i is the unknown phase for the
ith block, modeled as uniform over [0, 2], and Q() is the
quantizer.
AGC-Free ADCs: In order to further simplify receiver
design, we examine whether acceptable performance can be
obtained by removing the need for automatic gain control
(AGC). Specifically, by linearly combining the I and Q
components in the analog domain, followed by 1-bit ADC
(which requires no AGC), we can implement phase-only
quantization. An example is the 8-sector phase quantizer
shown in Figure 4, where we employ four 1-bit ADCs: two
for I and Q, and two for the linear combinations I+Q and I-Q
corresponding to a /4 phase rotation. Intuitively, we expect
phase-only quantization to work well with phase shift keyed
(PSK) constellations, but there are potential pitfalls, as we
illustrate next for QPSK and 8-sector phase quantization.
Information-theoretic results for phase-only quantization
are reported in [6]. Here we report uncoded performance of
24
(b)
t = 4dT
Fig. 5. (a) Standard PSK : the same constellation (the one shown) is used
for both symbols in the block. (b) Dithered-PSK : the constellations used for
the two symbols are not identical, but the second one is a dithered version of
the first one.
ADC
t = (4d+1)T
ADC
Analog
Input
0
10
Digital
Ouput
ADC
10
t = (4d+3)T
12sector phase
quantized ; No Dither ;
L=2,4,6,8
(dashed curves)
10
4-to-1
MUX
t = (4d+2)T
10
ADC
8sector quantized
Dithered; L=2,4,6,8
(dashed curves)
2
Fig. 7.
10
10
Coherent
QPSK
Coherent
QPSK
10
10
Unquantized
L=2,4,6,8
(dashdot curves)
10
10
12
Es/No (dB)
Unquantized
L=2,4,6,8
(dashdot curves)
10
14
16
10
12
14
16
Es/No (dB)
Fig. 6.
Symbol error rate performance for QPSK with 8-sector phase
quantization (left figure) and 12-sector phase quantization (right figure), for
block lengths varying from 2 to 8. Also shown for comparison are the curves
for coherent QPSK, and noncoherent unquantized QPSK.
A. System model
To a first order approximation, the mismatch between the
parallel branches is assumed to be gain, timing and voltageoffset mismatches [9]. Denoting the analog input to the ADC
as r(t), the digital output of the TI-ADC with L sub-ADCs is
given as,
r[m] = gmmodL r(mTo + mmodL ) + mmodL + q[m]
(3)
25
10 subADC
8 subADC
10
r = A(H B) + n
20
30
40
50
60
70
80
90
20
40
60
80
100
120
Interfering subcarrier
Fig. 8.
Effect of TI-ADC mismatch on the OFDM system: Relative
interference seen at the frequency bin corresponding to the first subcarrier in
128-subcarrier OFDM system. The values in the legend indicate the number
of sub-ADCs interleaved in the TI-ADC. Gain and timing mismatches are
uniform in the range 10% from their ideal values. [12]
interference power at the first sub-carrier due to other subcarriers. In general, all subcarriers interfere with each other, but
when L, the number of sub-ADCs, divides M , the number
of subcarriers, each subcarrier incurs interference from, and
interferes with, precisely L 1 other subcarriers. We therefore
assume that L divides M , in which case the subcarriers fall
into K = M
L mutually disjoint interference groups of size L,
so that mismatch compensation can be applied separately to
each group. For the j th interference group, j = 0, 1, ..., M
L 1,
we obtain a multiuser model as follows:
Rj =
L1
v 1[ y ]
1
FFT
(4)
r[m]
i=0
D[m]
IFFT
FFT
u *2 [ m ]
*
H[y]
2
|H[y]|
B[y]
*
H[y]
2
|H[y]|
B[y]
v 2[ y ]
(a)
u *1[ m ]
V1[ m ]
FFT
r[m]
D[m]
V2[ m ]
u *2 [ m ]
(b)
Fig. 10.
{(a) Double FFT-based implementation (b) Single FFT based
implementation } for time-frequency multiplier based equalizer to compensate
TI-ADC mismatch in an OFDM system. Here, D[m] is chosen appropriately
so as to cancel all the mismatch-induced interference [13]. In (b), cyclic
convolution is performed with the filters indicated by Vi [m] which are
obtained by taking IFFT of the vectors vi [y].
u1[ m ]
IFFT
(6)
Bj+iK Uj+iK + Nj
(5)
FFT
b[m]
B[y]
r[m]
H[y]
v*2 [ y ]
u2 [ m ]
Fig. 9. Approximate model for a TI-ADC, with gain and timing mismatches,
in an OFDM system
26
10
10
10
V. C ONCLUSIONS
BER
BER
10
10
10
Mismatch uncorrected
Conventional correction
Proposed (IV.B)
No mismatch
10
Mismatch uncorrected
Proposed (IV.C)
No mismatch
Conventional correction
10
10
10
0
10
15
20
10
15
20
R EFERENCES
D. Numerical Results
We now apply the proposed OFDM-specific mismatch compensation algorithms to an OFDM system employing 16-QAM
over a dispersive channel. Two values of interleaving factor,
moderate (L=8) and large (L = 32), are considered. In both
cases, 10% gain and timing mismatches are assumed between
the interleaved ADCs. See [12], [13] for further details. We
compare the performance of the proposed schemes with a
conventional mismatch compensation approach of employing
a zero-forcing (time varying) equalizer on the time domain
output of the TI-ADC [10].
Figure 11(a) depicts the BER performance before and after
mismatch correction for 8 sub-ADCs. It can be seen that the
method in Section IV.B is effective in eliminating mismatchinduced error floors, and outperforms conventional mismatch
correction. The number of taps in the conventional equalizer
was chosen to be L(=8), so as to match the complexity of
the proposed scheme. Figure 11(b) depicts the performance
for 32 sub-ADCs with an aggregate sampling rate equal to
twice the Nyquist rate. Using the time-frequency multiplier
scheme in Fig. 10(b) (Section IV.C), with the number of taps
in V1 [m] and V2 [m] as 1 and 2, respectively, we see that
the degradation due to mismatch is eliminated. In this case,
however, the conventional time-varying equalizer with 5 taps
for each sub-ADC also works well, matching the performance
of the time-frequency multiplier approach. This contrasts with
the performance of the (symbol rate sampled) conventional
scheme in Figure 11(a), and indicates that an oversampled
conventional scheme is competitive with the time-frequency
multiplier scheme. The time-frequency multiplier requires
an FFT of twice the length, but has a smaller number of
taps. In our example, the conventional scheme requires 5L
taps, five each for the L time-varying filters corresponding
to each sub-ADC, with the number of taps depending on
the desired resolution. For the time-frequency multiplier
approach, we need 2L taps for the two multipliers: u1 [m]
and u2 [m] (periodic with period L from [13]), and 3 taps for
the filters: V1 [m] and V2 [m], with a total of 2L + 3 taps. The
27