Signal Processing - September 2017 - 161

nonlinear cancelation schemes are preferred, but the MMSE
However, the major complexity occurs when the N # N
linear scheme still gives reasonable performance, with at most
canceler/precoder matrix is applied on the received/transmit
a 5% degradation compared to the MFB and 3% compared
vector of size N. This requires KN 2 fs instructions per second,
to ZF-GDFE. The nonlinear schemes are computationally
which becomes prohibitively high, at 1.9 billion instructions
more complex than the linear techniques. However, their
per second for an N = 10 users binder operating at fs = 48 kHz
performance can be improved with proper ordering procefor 4,096 tones.
dures among users [66]. Recent works [80] show that the
impact of the channel estimation error on the nonlinear
G.fast deployment and coexistence
is more crucial on the G.fast system than the VDSL sysWith the penetration of fibers closer to the end users, various
tem. This is also an important design consideration for the
deployment scenarios have been considered for the G.fast
nonlinear G.fast system.
FTTdp system. The FTTdp consists of sceIt is noted that the above simulation
narios where the DP is either mounted on a
Compared to wireless
results were obtained without optimal popole or placed underground at the curb,
communications, the
wer control (as discussed in the section
building, basement, and so on. The G.fast
G.fast requires more
"Joint Multiuser Crosstalk Cancelation and
FTTdp fiber node can be as far away as 250
Dynamic Spectrum Management"). Algom to serve dozen of users and as close by as
complicated processing,
rithms for joint DSM with crosstalk can50 m to serve a single user. Because such a
as it simultaneously
celation in [67] and [68] showed that there
scenario requires the deployment of many
serves more users and
was a significant performance improvement
new DPs, the G.fast also considers the poswith higher user rates.
for the optimized linear ZF and nonlinear
sibility that the DP will be powered from
ZF-GDFE at higher frequencies. Howthe CPEs. This technology, known as
ever, the considered algorithms had higher computational
reverse powering, eliminates the need for a power infrastruccomplexity because of the large-dimensional power allocation
ture for the DP and hence lowers the deployment cost. It also
in G.fast systems.
introduces new challenges, such as the use of a single wire for
both broadband and power transmission and issues related to
Implementation complexity
safety guidelines for power transmission in a home network.
Crosstalk cancelation schemes require significant computational
Another issue affecting G.fast deployment is the overlap of
resources, given various signal processing operations at a large
its bandwidth with other radio and broadcast services. Radio
scale. The first is the memory requirement of the processor to
services overlap a few consecutive tones with the G.fast system
store N 2 channel coefficients per tone. Because the number of
throughout the band. These tones need to be either masked or
tones is very large, K = 4,096, the memory requirement is signotched, such that no interference is experienced by the radio
nificant. The efficient use of memory for this task also needs
service. Broadcast radios such as FM are broadband services
further investigation. For example, the memory requirement can
and cover a large bandwidth (20 MHz). The G.fast system has
be reduced if the channel coefficients are stored at a few tones
to avoid data transmission over these bands, which limits the
throughout the band and interpolated on the other tones.
rate performance of G.fast. Note that these unused tones can
However, this will reduce the data rate performance, because
be employed for system provisioning. It should be emphasized
the error in channel estimates increases with the interpolation
that efficient transmissions methods are used to enable the
interval. These tradeoffs are standard signal processing design.
coexistence of the G.fast system with other broadband techThe loss due to quantization and interpolation can be characternologies without disturbing the band profile.
ized using the techniques in [81].
The coexistence of the G.fast system with legacy lines is
Another complexity consideration is the computation of the
another important issue [82]. Vectored VDSL systems have
canceler/precoder matrix. The linear ZF and MMSE canceler/
already been deployed by various service providers and may
precoders require inversion of many channel matrices, the
coexist with G.fast users in the same binder. VDSL fallback
complexity of which is very large O (KN 3). The QR decompoand coexistence are necessary for mass deployment of G.fast.
sition in the ZF-GDFE canceler/THP precoder requires simiMost service providers require that migration from VDSL to
lar high computational resources. The approximate inversion
G.fast should be seamless, such that the G.fast must be able
method [12], which avoids the matrix inversion, is not recomto peacefully coexist with VDSL2 from the same DP. A commended for the G.fast system at higher frequency tones.
mon approach to coexistence is to have both technologies use
The adaptation of the canceler/precoder coefficients should
nonoverlapping spectra. This is a simple approach to avoid any
be simple and not need costly computation. To track the chaninterference between the two technologies. However, leaving
nel, the G.fast standard provisions a single sync symbol after
the lower frequency bands where the spectral efficiency is very
every 274 data symbols. The authors in [61] proposed a scheme
high limits the data rate performance of the G.fast technology.
to directly track the canceler matrix, whose convergence
The coexistence of G.fast and VDSL in the same DP with an
depends on the diagonal-dominance parameter of the channel.
overlapping spectrum is a challenging problem, somewhat
Further research is required to derive computationally efficient
similar to the problems faced by ADSL deployment in Japan
adaptive canceler schemes for the G.fast channel.
parallel to TDD-based high-bit-rate DSL [83].
IEEE SIGNAL PROCESSING MAGAZINE

September 2017

161

Table of Contents for the Digital Edition of Signal Processing - September 2017