ABSTRACT
`The Internet has surfaced as the dominant early market for residential broadband. ADSL, a transmission system capable of realizing rates
`from 1 t o Mb/s over existing telephone lines, fits Internet access requirements perfectly, and offers telephone companies a tool for
`connecting virtually all Internet users at megabit rates before the next century. ADSL is asymmetric - high-speed downstream,
`lower-speed upstream - to counteract speed limitations imposed by line length and crosstalk. The transmission technology itself has
`two essential forms, single-carrier and multicarrier, which must press Shannon‘s limit t o squeeze so many bits through so little
`bandwidth. With complicated line coding and other features such as integral forward error correction and ATMiEthernet mode
`interfaces, ADSL will be the most complex modem ever attached t o a telephone line. This will not prevent ADSL from reaching
`consumer-level pricing within the next t w o years. We can expect some commercial deployment in 1997
`and virtually ubiquitous availability by the end of 1999.
`
`Kim Maxwell, Independent Editions
`
`fter much backstage preparation, asymmetric digital
`subscriber line (ADSL) is about to burst through the
`curtain. It will take a short bow, wait for a few protocol props
`left behind (financing came too late to get a full dress
`rehearsal), and stumble around for awhile as editors put in
`last-minute changes. But then ADSL will do its act, to multi-
`ply by the millions, and carry megabit data to users around
`the world over their existing plain old telephone lines.
`This article offers a glimpse of ADSL as it is today, just
`about ready for that first bow. It will be broad, a brief techni-
`cal taste of the ADSL world, inside and outside the modem.
`Readers frustrated by the consequent failure of depth can find
`all they need in the reference works on some subjects, but
`some important parts of the ADSL act still await final lines.
`For this we should be glad. We are innovators, after all, not
`historians.
`
`WHY SHOULD WE CARE?
`I fibericoax. Seen as a competitive response to imminent com-
`n 1993 many telephone companies were agog about
`
`petition from cable television (CATV), HFC would hit the
`streets running in 1994 and be in full deployment - many
`millions of lines a year - by 1996.
`It is now 1996, and HFC has virtually died as a strategic
`component of next-generation access networks. HFC costs
`grew and grew (and still grow), and HFC bandwidth cannot
`support a full range of services to all customers. Furthermore,
`no single idea has taken its place, and telephone companies
`seem far more committed to service-specific networks than
`any near-term rollout of a full-service network. The reason is
`not obscure. The only new broadband application with univer-
`sal appeal is video on demand. With today’s technology, video
`on demand cannot generate enough prospective revenue to
`justify new infrastructure, and this equation is likely to hold
`for several more years. Therefore, telephone companies con-
`sider, and plan to deploy, MMDS, LMDS, fiber to the curb
`(FTTC), fiber to the node (FTTN), fiber to the home (FTTH),
`some WFC (often with purchased CATV companies), SDV,
`
`and ADSL, each to specific market segments; they hope that
`the various fiber alternatives can migrate to a full-service net-
`work over time.
`ADSL would not have been on this list three years ago.
`ADSL runs at megabit rates - up to 9 Mbls downstream and
`up to 1 Mb/s upstream - over existing copper telephone
`lines. Copper was scorned three years ago, a retrograde, rear-
`view-mirror technology with limited capacity and hopefully
`limited life, to be shuffled off this mortal coil without ceremo-
`ny. That copper connects 700 million locations (and serves
`well over a billion users), and constitutes the last mile and
`largest single expense for an industry ringing up $750 billion
`in annual sales, seemed as irrelevant as wolves baying at the
`moon.
`What happened?
`The Internet. To the surprise of many, the Internet has
`rocketed into such prominence that simply uttering the word
`before securities analysts doubles a company’s stock price. In
`simple terms, the Internet is a widely dispersed packet-mode
`cloud suffering, now, from too many subscribers, too little
`backbone bandwidth, slow routers, low server bandwidth, and
`miserable access speeds. This is a perfect context for ADSL.
`Assuming the Internet itself grows suitably (and it must),
`ADSL instantly increases access speeds by two orders of mag-
`nitude, to rates likely to be faster than the Internet itself can
`support for a number of years. Copper already connects all
`Internet users. Once ADSL access networks reach maturity
`(by 1998), ADSL can be deployed so fast that virtually all U.S.
`prospects can be connected by the year 2000. And the most
`important feature of ADSL access is that realistic revenue
`projections exceed costs from the outset. Indeed, it is not hard
`to make a strong case for ADSL serving tens of millions of
`customers and being around as long as copper, that is, for
`decades. As Ray Smith of Bell Atlantic opined, “ADSL is an
`interim technology, for the next forty years.”
`ADSL, however, does not stand for All Data Subscribers
`Living - it will not work over every telephone line, and cer-
`tain telephone line parameters limit ADSL rate and perfor-
`mance. Grasping ADSL technically requires some
`
`100
`
`0163-6804/96/$05.00 0 1996 IEEE
`
`IEEE Communications Magazine October 1996
`
`CSCO-1015
`Cisco v. TQ Delta
`Page 1 of 7
`
`

`
`understanding of these parameters.
`So this essay will begin with what
`exists, the telephone plant.
`
`THE TELEPHONE PLANT
`T Dlant of teleuhone comDanies
`he so-called subscriber loop
`
`today consists primarily of unshield-
`ed twisted pa& copper access lines
`with passive premises terminations. In some countries, notably
`Germany, as much as 5 percent of the loop pPlant is digital,
`with active integrated services digital network (ISDN) termi-
`nations. In the United States (and some other countries),
`approximately 15 percent of the loop plant terminates the net-
`work side of copper lines in remote digital loop carrier (DLC)
`systems which multiplex voice lines over copper or fiber from
`an outside plant location to a central office. Of the 700 mil-
`lion lines operating today, about 70 percent serve residences,
`with the balance serving businesses. The United States today
`has about 160 million access lines.
`Twisted pair copper attenuates signals proportional to
`length and frequency. If lines get sufficiently long (about
`15,000 ft for 26 gauge wire, 18,000 ft for 24 gauge wire), their
`cumulative dc impedance begins to affect voice quality and dc
`signaling reliability. To compensate (in the United States. at
`least), telephone companies install loading coils in the line
`that effectively filter all frequencies above 4 kHz, and thereby
`bar any DSL service, including ISDN. Somewhere between 15
`and 20 percent of all U.S. residential lines have loading coils.
`Long lines also attenuate across the band; the canonical
`18,000-ft line has a 50 dB slope over the normal band of fre-
`quencies used for ADSL downstream data.
`The figure of 18,000 ft has become a frequently cited nor-
`mative bound for ADSL and ISDN, but it is a loose and ficti-
`tious one. It applies only to continuous runs of 24-gauge wire
`without bridged taps. There are almost no such lines in prac-
`tice. Telephone companies pull 26-gauge wire, in bundles of
`1000 or so lines, from central offices, and convert to 24-gauge
`about 10,000 ft out to improve impedance versus distance;
`rural areas may even see 19-gauge wire. DLC sites may
`encounter 24-gauge wire directly, but they seldom support
`lines longer than 9000 ft. Furthermore, plant cabling tends to
`come in 500-ft lengths, meaning a splice every 500 ft. Bellcore
`estimates that the average line has 22 splices; splice points
`collect corrosion and add attenuation if poorly made. Finally,
`many U.S. lines have bridged taps, a second (or third or
`fourth) unterminated spur off a line that may be quite short
`or thousands of feet long. Each bridged tap acts like a delay
`line and puts a notch in a line’s frequencylattenuation charac-
`teristic at the frequency associated with a bridged tap’s wave-
`length.
`Attenuation dominates the factors limiting ADSL perfor-
`mance, but two other parameters have important effects:
`crosstalk and impulse noise. Alexander Graham Bell invented
`twisted pair wiring to, among other things, minimize coupling
`of signals from one pair to an adjacent pair when lines were
`bound together in a cable. The process is not perfect. Signals
`do crosstalk from one pair to another, at levels that increase
`with frequency and the number of crosstalking pairs, or dis-
`turbers. (The model used in ADSL standard T1.413 shows
`crosstalk increases proportional to frequency raised to the
`power 312 and to the number of disturbers raised to the power
`0.6. Note, however, that new cables, such as UTP Category 5,
`improve crosstalk performance by as much as 20 dB over
`existing installed telephone wire.) As noted above, line attenu-
`ation also increases with distance and frequency. These fac-
`
`ADSL does not stand for
`All Data Subscribers Living;
`it will not work over every
`telephone line, and certain
`telephone line parameterS limit
`ADSL rate and performance.
`
`tors drive the “asymmetric” nature
`of ADSL.
`If a modem can realize 6 Mbls
`on a given line, it can do so in both
`directions at the same time with
`suitable echo cancellation (the tech-
`nique used in V.32 and V.34 to sep-
`arate umtream from downstream
`channels). However, putting two
`such devices in the same cab6 will
`likely bring both to a halt. At both ends the adjacent transmit
`signal crosstalking into a local line above a certain frequency
`will essentially destroy the weakened local receive signal. This
`frequency, of course, depends on line length and gauge and
`the signaling complexity of the modem itself. But high-bit-rate
`DSL (HDSL), a symmetric service, transmits a duplex signal
`of no more than 750 kb/s (in a band of 240 kHz) for distances
`of 12,000 feet of 24 gauge wire (HDSL uses two lines and
`inverse multiplexing to achieve 1.5 Mbls). Crosstalk prevents
`higher duplex rates with HDSL‘s line code.
`ADSL beats this problem by sending in one direction only
`- downstream -with a much lower upstream rate separated
`from the downstream by frequency division multiplexing
`(some echo cancellation is possible at low frequencies). Cur-
`rent ADSL products use 25 to 250 kHz for the upstream. and
`25 kHz to above one MHz for the upstream. As we shall see,
`the upper limit depends on the data rate and modulation sys-
`tem used. Note that an inverse ADSL with a high-speed chan-
`nel going upstream (e.g., for an Internet server) must be
`disallowed. It will work, but it will either slow down or stop
`any other ADSL modems in the same cable with the conven-
`tional configuration.
`Attenuation and crosstalk normally make up the canonical
`impairments for defining DSL performance. With crosstalk
`representing reasonable fill rates of a cable, the following
`downstream rates can be realized for the indicated distances
`of 24-gauge wire:
`18,000 ft
`1.5 Mbls
`2.0 Mbls
`16,000 ft
`12,000 ft
`6.0 Mbls
`9000 ft
`9.0 Mbls
`13.0 Mbls
`4500 ft
`3000 ft
`26.0 Mbls
`1000 ft
`52.0 Mbls
`The last three rates fall under VDSL rather than ADSL.
`As suggested above, 18,000 ft encompasses about 80 percent
`of lines in the United States. The region called the carrier
`serving area extends to 12,000 ft, and encompasses about 50
`percent of lines in the United States. The faster rates on
`shorter loops will almost certainly be implemented in outside
`plant in various forms of fiber to the .... In addition, tele-
`phone companies will reach subscribers who fall outside the
`range of ADSL reach by installing fiber-based concentrator
`nodes, which will be stepping stones to deeper penetration of
`fiber into the loop plant.
`Crosstalk noise is usually stationary. Impulse noise is ran-
`dom, in frequency, duration, and amplitude. As a result, it is
`difficult to model or study empirically. Furthermore, the
`impulse noise that arises in the telephone system has tolerable
`effects on voice communications and data communications
`using the 4 kHz voice channel available to ordinary modems.
`Thus, there has been little incentive to measure it or model it
`until recently. The picture emerging from the few field surveys
`published suggests that, while many impulses are small and
`short, a significant number, particularly those arising from
`ringing and trip ringing in adjacent pairs, can have destructive
`
`IEEE Communications Magazine October 1996
`
`101
`
`Page 2 of 7
`
`

`
`network - I
`
`Private
`
`Public
`network
`
`
`
`-
`
`1
`
`ADSL access network
`1-9 Mb/s pb
`
`~
`
`I
`
`Premises
`network
`
`-4
`1
`-4
`1
`
`1
`
`ATM
`ATM
`
`Packet
`
`4
`
`STM
`Packet
`
`b - 4
`b - 4
`ATM
`
`ATM
`
`Transport modes
`
`Figure 1. ADSL network diagram.
`
`amplitudes for more than 1 ms. The DSL industry has a sort
`of working model now that shows 75 percent of all impulses
`with impulse width below 500 ps, and a pulse shape defined in
`T1.413, called a “Cook” pulse, that can easily be simulated.
`However, the real impulse world remains rather mysterious.
`
`GETTING ORIENTED
`efore launching into ADSL details, it will be worth taking
`one or two pictures from a few thousand feet up. Figure 1
`shows the ADSL Forum Network Model. In essence, ADSL
`uses existing telephone lines to connect user terminals - per-
`sonal computers and televisions - to various services over
`tandem combinations of public and private networks at much
`higher speeds than can be realized today with voice-band
`modems or ISDN. The public network part comprises an
`access node, for concentration and perhaps protocol conver-
`sion, and a switching or routing fabric. Access nodes may be
`located at central offices or in the loop plant at the end of a
`fiber link. Switching or routing facilities may be at central
`offices or buried deeper in the network. A controversy storms
`
`“a
`
`U-c2 U-c
`
`U-R U-R2
`
`T-SM
`
`T
`
`Figure 2. ADSL system reference model.
`
`ADSL
`
`STM
`Packet
`
`b-
`
`network
`
`PDN=Premrses distribution
`b
`b
`b
`b
`
`Packet
`
`b
`
`today about transport protocols - will EthernetiIP or ATM
`dominate access node multiplexing? The pendulum seems to
`be swinging toward ATM, in which case the switching point
`will be in higher-level offices supporting numerous end offices
`and remote access nodes over fiber.
`While the network side of this picture is complex, it has far
`fewer clouds hanging over it than the customer premises.
`These clouds linger above the innocent box called “premises
`distribution network,” which can be anything from simple
`wiring to an Ethernet LAN to, sometime in the future, an
`ATM network connected to a residential gateway. It is not
`that any particular configuration stumps experts; it is that
`there are so many of them, all subject to the liabilities of cus-
`tomer installation. Unless an industry is developed to install
`premises networks, the inexpert user will be faced with piecing
`together ADSL with plain old telephone service (POTS) split-
`ters, wiring, personal computers with or without network
`information center (NIC) cards, NIC cards, hubs, perhaps
`routers, and various software packages to pull usable data
`from whatever format ends up at the computer interface.
`(Prediction: such an industry will develop.)
`Assuming telephone companies adopt a homogeneous
`ATM network to the premises, the most likely transport
`mode for the next few years will be the next to last shown in
`Fig. 1, with a large packet interface (such as frame relay)
`between the service provider and the public network and
`Ethernet between the ATU-R and the personal computer.
`The latter may use a Cells-in-Frame (CIF) protocol to tunnel
`ATM through a premises Ethernet. In any event, modems at
`both ends must be ATM-aware (cell pumps rather than bit
`pumps), and perhaps include some protocol conversion at the
`premises.
`Figure 2 shows part of the ADSL Forum System Refer-
`ence Model (a subset of the network model). At its center is
`the only thing that really exists in volume today, the tele-
`phone line. As ADSL shares this line with POTS, the first
`thing the line encounters on each end is a set of filters that
`
`102
`
`IEEE Communications Magazine October 1996
`
`Page 3 of 7
`
`

`
`Line
`interface
`-
`
`‘i ~i~~~~ 3.ADsL functional layers.
`
`Digital
`interface
`
`split the line by frequency, a low-pass filter
`passing POTS and a high-pass filter passing
`ADSL signals at roughly 25 kHz and above.
`The POTS line goes off to telephones at
`the customer premises, and the public
`switched telephone network (PSTN) at the
`network end. How to accomplish this and
`not disturb either the quality or reliability
`of POTS is neither trivial nor completely
`solved today. The largest question mark
`concerns the U-R2 and U-C2 interfaces.
`The ADSL Forum has recommended, and
`TlE1.4 has agreed. that the POTS snlitter
`should be phYysicaily separated f r i m the
`modem (the roughly 10,000 ADSL units in
`the field‘today iitegrate the POTS splitter
`with the modem). This raises the rather difficult problem of
`defining an interface so a POTS splitter can be purchased
`from one vendor and a modem from another.
`When Bellcore first conceived ADSL (1989), they envi-
`sioned a simple bit pump with a 1.5 Mbls downstream rate
`and a 16 kbls or 64 kb/s duplex channel for signaling and
`video controls, targeted at video-on-demand applications.
`Today some ADSL modems realize downstream rates of 9
`Mbls, upstream rates of 1 Mbls, initialization protocols that
`will pick the best speed for a given line, and packet or cell
`interfaces that connect directly to Ethernet or ATM premises
`distribution networks. The simple bit pump is probably still-
`born. Over its grave stands a suite of features sufficiently com-
`plex that standards groups are now considering a division of
`the basic modem into two layers - the physical-media-depen-
`dent (PMD) layer and the transmission convergence (TC)
`layer, following a similar division in ATM physical (PHY)
`layer protocols. Figure 3 suggests the divisions of functions for
`each layer and the various versions of each layer’s implemen-
`tation that might be considered.
`
`standardized line code matures into viable commercial prod-
`ucts. Both use single carrier techniques - auadrature ampli-
`tude modulation (QAM), the mother of all ADSL line codes,
`and carrierless amplitude-phase modulation, AM-PM (CAP),
`a variant of QAM with some implementation and flexibility
`benefits over QAM; these are discussed in some detail below.
`One normally begins a discussion of modems with rates
`and bandwidth: what rate does the line code have to realize
`over what frequency range? ADSL began in this conventional
`manner - downstream rates of 1.5 Mbls over 18,000 ft of 24-
`gauge wire, 6 Mbls over 12,000 ft of 24-gauge wire, assuming
`certain models of crosstalk interferers - but early in 1996
`ADSL took an odd turn. Someone observed that ADSL was
`fishing for business in the Internet lake rather than the video
`sea. The Internet is inherently variable-rate, promises of real-
`time services notwithstanding. Interfaces to the network and
`to the home PC (not the television settop box) would be Eth-
`ernet or ATM, both variable-rate. Why not make ADSL vari-
`able-rate, offering the subscriber the best rate his line would
`allow, even if this rate fell below 1.5 Mbls? The Internet
`wasn’t going to run even that fast for most uses anyway for
`quite some time.
`Now this idea has the blessed property of extending the
`number of telephone lines ADSL will work on without line
`engineering. It was consequently endorsed by telephone com-
`panies with remarkable rapidity. To avoid the image of ADSL
`adapting continuously to small line variations, “variable rate”
`was changed to “rate adaptive,” giving rise to a new acronym,
`RADSL, for rate-adaptive DSL. RADSL modems will likely
`dominate the near-term market, particularly in the United
`States. Some countries still pursuing ADSL for video delivery
`will stay with fixed-rate ADSL. In practice, the two modems
`will also likely be the same, since fixed-rate ADSL can clearly
`be carved out of RADSL, and RADSL is, in practical terms,
`no more expensive than ADSL.
`Rate adaptation is not restricted to the downstream chan-
`nel. Some studies show that good
`performance on the Internet
`;quires a downstream/upstream
`ratio of 1013. While protocol
`tweaking and parameter negotia-
`tion can raise this ratio to, say,
`2011, the former figure has
`become an operating target. Thus,
`a downstream rate of 2 Mbls
`needs an upstream rate of 200
`kbls, while a downstream rate of
`6 Mbls needs one of 600 kbls.
`This sort of rate flexibility also
`extends to other services. For
`
`PHYSICAL- M EDI A- D E PE N DE NT
`ADSL FUNCTIONS
`T mally think of as a modem. Regardless of modulation
`he PMD section of ADSL represents what we would nor-
`
`technique, all ADSL transceivers perform the functions shown
`in Fig. 4. A modulator creates a digital representation of a
`signal modulated by the particular combination of transmit
`data bits during any given symbol period (inverse of the baud
`rate). An analog section converts this digital representation to
`analog, filters it, and then amplifies it to a level consistent
`with line power requirements. The receiver section essentially
`reverses this process, but must equalize the line to normalize
`the signal beforehand.
`At present ADSL has three candidate modulation tech-
`niques, or line codes, making
`their way to the marketplace.
`One, discrete multitone (DMT),
`divides the line into many small
`channels and modulates each one
`based on its capacity for a given
`line; ANSI standards group TlE1.4
`has developed an ADSL standard,
`number T1.413, around DMT.
`However, two major telecommu-
`nications suppliers have embarked
`on alternative line code imple-
`mentations in an effort to seize
`early market share before the
`
`Figure 4. Basic tranvxzver PMD layer.
`
`IEEE Communications Magazine October 1996
`
`103
`
`Page 4 of 7
`
`

`
`POTS
`
`Downstream
`
`Downstream
`
`1 MHz
`
`1
`
`but timing recovery in the receiver often dictates
`some excess bandwidth, usually on the order of
`15 percent. Thus, a 680 kbaud signal will use 782
`kHz of bandwidth.
`Using this bandwidth efficiently comes from
`QAM’s ability to assign increasing numbers of
`bits per symbol. A typical QAM system (say the
`one in V.32 for 9600 bis) groups four bits per
`symbol, and requires a 21 dB signal-to-noise ratio
`(SNR) at the receiver to realize suitable error
`rates. However, QAM can be designed with fewer
`bits per symbol, and with more, up to 15. For
`each bit added the SNR must increase 3 dB to
`achieve the same error rate, and each bit added
`pushes implementation constraints on noise floors, D/A and
`A/D converters, and DSP processor bandwidth. However,
`practical QAM implementations today achieve 8 bisymbol,
`meaning, for example, that a 680 kbaud output transmits 5.44
`Mbis, but requires 33 dB SNR at the receiver.
`Note: It is common to talk about signaling densities as bits
`per Hz, or bits per baud, or as the number of points in a
`constellation associated with a particular number of bits
`per symbol. Constellations usually have a number of points
`equal to 2 raised to the power of the bits per symbol. For
`example, 4 bisymbol is the same as 4 b/Hz and 16 QAM,
`and 8 b/Hz is the same as 256 QAM.
`The problem and complexity for any single-carrier system
`for ADSL arises from how to adapt to wide variations in tele-
`phone lines. To work well, a QAM receiver needs an input
`signal with the same spectral shape and phase relationships as
`the one transmitted. Telephone lines change both; therefore,
`QAM receivers include adaptive equalizers that determine
`line characteristics and use them to compensate for distortion
`added during transmission. The process is not perfect, and the
`wider the range of possible distortions, meaning the wider the
`range of lines, the more complex the equalizer must be.
`Indeed, for ADSL the adaptive equalizer dominates system
`complexity for QAM implementations.
`AT&T has developed a variation on QAM, called carrier-
`less AM-PM or CAP, which generates a transmit waveform by
`applying each half-rate bitstream to a pair of digital transver-
`sal baudpass filters with equal amplitudes but phase responses
`differing by x/2. This produces the same spectral shape as
`QAM, may be detected with the same equalization strategies,
`and has the same performance as QAM. Indeed, a QAM
`receiver can be modified to receive a CAP transmit signal.
`CAPS virtue lies in some efficiencies compared to QAM with
`digital implementation.
`The initial CAP systems for ADSL followed Bellcore’s lead
`and implemented a single downstream rate of 1.5 Mb/s and an
`upstream rate of 64 kbis. The latest proposed CAP system is
`rate-adaptive, from 640 kb/s to 8192 kbls in the downstream
`direction and from 272 kbls to 1088 kbls in the unstream
`direction. To achieve reasonable granularity, the proposed
`CAP system implements five downstream baud rates with five
`
`25 kHz 200 kHz
`
`~
`
`1 Frequency division multiplexing (FDM)
`
`1 MHz
`
`25 kHz 200 kHz
`Echo cancellation
`
`Figure 5. Channel configurations.
`
`example, a user may want a video conference at 384 kb/s sym-
`metric during one session and a very asymmetric movie need-
`ing 1.5 Mbls during another session. A rate-adaptive modem
`(particularly one with echo cancellation) can rearrange the
`two directions to suit (subject to the limits imposed by line
`attenuation and crosstalk PSD masks).
`Figure 5 shows channel allocation for two basic ADSL
`modes. Each mode blocks off the lower 25 kHz for POTS
`(POTS only needs 4 kHz, but POTS splitters become very dif-
`ficult to design if the lower edge of the upstream channel gets
`any closer). An upstream channel with usable bandwidth on
`the order of 135 kHz takes the next slot. This section of the
`channel has the most favorable attenuation characteristics, but
`also suffers the most crosstalk from other services such as
`ISDN DSL (with frequencies to 80 kHz) and HDSL (with fre-
`quencies to 240 kHz). In the rate-adaptive mode, oddly
`enough, the upstream channel may be the limiting resource
`rather than the downstream in some circumstances.
`In frequency-division multiplexing (FDM) the downstream
`channel starts above the upstream, at approximately 240 kHz,
`and extends as far up as needed, or permitted, by a combina-
`tion of desired data rate, attenuation, and modulation mode.
`As we will see below, a rate-adaptive single-carrier system can
`use bandwidths anywhere from 340 kHz to 1088 kHz to
`achieve data rates from 680 kbls to 8.7 Mbls, graduated in
`steps of about 320 kbis. A multicarrier modem, with its chan-
`nel slivers adapting to line conditions, may use frequencies
`out to 1.1 MHz (a band of 860 kHz): its rate range is from 32
`kbls to in excess of 9 Mbls, graduated in steps of 32 kb/s.
`In echo cancellation (EC) mode the downstream channel
`overlaps the upstream. This has two advantages, at the cost of
`the echo canceller: the downstream has more bandwidth in
`the good part of the line; and the upstream can be extended
`upward without running into the downstream. In practice the
`latter is the most significant benefit. At present only multicar-
`rier ADSL modems have been implemented with echo cancel-
`lation.
`
`SINGLE-CARRIER MODULATION
`The parent line code for all ADSL is QAM. Of the various
`general modulation schemes available, QAM has the best
`combination of bandwidth efficiency, perfor-
`mance in the presence of noise, and timing
`robustness. (Other line codes, such as 2B1Q
`used in ISDN DSL and HDSL, have virtues,
`largely in relatively low complexity and cost.)
`As suggested by Fig. 6, a QAM bitlsymbol
`encoder forms bit groups during each symbol
`period and then splits them into half rate
`streams that modulate a pair of orthogonal car-
`riers, which are in turn summed to form the
`output waveform. The output band must be at
`least equal to the symbol rate (the baud rate),
`
`104
`
`IEEE Communications Magazine * October 1996
`
`Page 5 of 7
`
`

`
`Bits per hertz
`
`Line gain
`
`Bits per hertz
`
`I
`
`Frequency
`.
`.
`
`
`
`Frequency
`.~
`
`.~
`Frequency
`
`different implied carriers and constella-
`tions ranging from 8 (3 bisymbol) to 256 (8
`bisymbol). The lowest baud rate uses spec-
`trum out to 631 kHz; the highest goes out
`to 1.491 MHz. The modem determines
`during initialization which combination
`Provides the best rate for
`Particular
`line and sets up the system accordingly.
`MULTICARRER MODULATION
`As the name implies, multicarrier modulation divides a chan-
`nel into numerous subchannels and transmits data on each
`one. The technique has a long history and considerable theo-
`retical support as an optimum code, but has been troubled by
`the cost of replicating transceivers and stability problems with
`analog circuits. In the early 1980s it was shown that multiple
`channels could be realized with digital techniques using a fast
`Fourier transform (FFT), giving rise to DMT, the version of
`multicarrier used in ADSL. A DMT transmitter encodes bits
`per symbol and loads them into an inverse FFT, pushing the
`output to a DIA converter. The receiver reverses the process,
`taking the output of the AID into an FFT and reassembling
`the serial bitstream from the result. Advances in signal pro-
`cessing technology now make FFTs practical for commercial
`products, and work over the last three years suggests that
`DMT, CAP, and QAM have comparable levels of complexity
`for ADSL (if anything, DMT enjoys a complexity edge because
`equalizers grow faster than FFTs when the going gets tough).
`Present ADSL designs with DMT create 256 downstream
`channels of 4 kHz with a aggregate symbol period of 250 ks.
`Each channel can be modulated (with QAM) at up to 15
`b/Hz. Theoretically, therefore, DMT could transmit 15.36
`MbIs over a line of zero length. Real lines and real implemen-
`tations, of course, are not so forgiving; but rather than use
`adaptive equalizers to compensate for variations in line atten-
`uation, DMT spreads data over all channels according to the
`SNR in each one. Figure 7 shows the adaptation process.
`During initialization a DMT modem measures the SNR per
`channel. With ADSL, of course, the POTS splitter essentially
`removes all signals below 25 kHz. An FDM arrangement
`would also remove the bandwidth needed for the upstream. At
`the upper end of the band attenuation can be severe, and
`bridged taps can create nulls in the attenuation curve. For a
`given data rate, then, DMT groups data in relatively large
`chunks (1500 bits for 6 Mbis) and spreads them over usable
`channels such that the margin in each channel is the same. In
`Fig. 7, for example, the first few channels have no data, the next
`few might have 2 or 3 biHz, the sweet spot may actually have
`as many as 15 b/Hz, and the trailing edge, at high frequencies,
`could drift back down to a low number of bits per hertz.
`DMT makes optimum use of the line by making optimum
`use of each subchannel. The channels are sufficiently narrow
`that they need little extra equalization (the attenuation slope
`is so bad they do need some). In a fixed-rate mode, DMT
`loads each channel such that the aggregate rate equals the
`input rate unless the margin per channel falls below a preset
`threshold, in which case the modem will not work. This means
`that most lines will have much higher margins than this
`threshold point. In the rate-adaptive mode, however, the
`modem picks the data rate based on how many bits can be put
`in each channel at thc preset margin ( a nc;twork operating sys-
`tem can set the rate lower if it chooses). Rate adaptation is
`quite simple for DMT. The initialization protocol must deter-
`mine margin in each channel anyway. It only takes a little
`framing adjustment to fit a computed data rate to a real data
`rate. As a consequence, DMT ADSL can rate-adapt in 32 kb/s
`increments, and go arbitrarily low - to 32 kb/s if necessary.
`
`Figure 7. DMT bitsper channel allocation.
`
`THE HATFIELDS AND MCCOYS
`So which is it, CAPIQAM or DMT? Shouldn’t there be just one,
`to make an efficient market and ensure interoperability among
`all ADSL vendors? Isn’t there a standard? What gives?
`There is a standard, T1.413, describing an ADSL system
`using DMT. There are also some