Rupert Baines
Analog Devices
TechOnline

With the internet explosion of the early 1990s, the global demand for high speed data delivery began to grow exponentially, in many different directions at once, virtually out of control. As we approach the new millennium, that growth seems to be reaching critical mass, and shows no signs of abating. Meanwhile, the present decade has presented the broadband industry with a staggering array of exciting opportunities.

So it seems almost a tragedy that we're still discussing the issue of DMT versus CAP. PHY layer discussions and line code squabbles are counterproductive given the many challenges we face in trying to commercialize broadband and successfully deliver high speed data access to millions of consumers worldwide. There are so many more important issues facing the broadband industry-billing, service management, backbone capacity and infrastructure, service architecture-why divert attention, effort, and resources away from them with unnecessary debate?

Unfortunately, the focus paid to this basic technology issue has perpetuated the mistaken belief that ADSL is an immature, laboratory technology not yet suited to mass deployment. Some of these claims have spread unfounded concerns about the suitability of ADSL as a broadband solution for operators (comments regarding power consumption or internet access suitability, for example).

Choice of line-code (modulation) method is an engineering decision, not one to be made on faith or gut instinct. Different situations suggest different choices. Certainly there are applications where single-carrier technology is well suited, and should be considered. But as we'll see in the next few pages, ADSL access to broadband services (internet access, LAN access as well videoconferencing, tele-learning, and video-on-demand) over the copper network is far better served by DMT.

This is not just our opinion; it is the consensus of experts around the world, global standards organizations that include ANSI, ETSI, and the ITU. And it's the opinion of many independent manufacturers who have the experience and expertise to design products using whatever code they deem best (they all selected DMT). Table 1 addresses some of the myths that have grown up around the debate, and answers each with what we believe to be the reality of the issue.

Table 1:  The myths and realities of CAP versus DMT

The two techniques can be seen as not-so-identical twins: CAP operates in the time domain; DMT, in the frequency domain. QAM/CAP techniques are time-domain based with fast symbols. Each symbol (and there's only one) lasts a short time, and has a sizable bandwidth. Each DMT subchannel (hundreds of them are communicated in parallel) lasts a long time but occupies a narrow frequency band. A DMT symbol, which is the combination of all the subchannels, lasts 250 ms—only a 4 kbaud symbol rate.

In principle, on a given channel, the two should achieve the same throughput, (Shannon's law does not specify line code). In practice, however, differences in the transmitter and receiver architecture as well as implementation limitations due to cost affect real-world performance.

The ADSL environment is different: high frequencies are attenuated more significantly than low frequencies; bridge taps can cause frequency notches and resonances; radio stations cause strong (but stable) RFI. It's a difficult channel in that it treats dissimilar frequencies very differently-but the channel properties don't change with time (or change very slowly). A DMT transmitter can easily monitor the channel, adapt its transmission to the characteristics of the phone line, and continuously update (via bit-swapping) to maintain the optimum. For every line, the DMT system transmits the "best" possible signal. A CAP system can't modify its transmitter, so it tries to undo all the attenuation and notches in one fell swoop at the receiver-a tough challenge. That's why, in evaluations, CAP systems are often described as being less robust than DMT systems, or not as tolerant to bridge taps. DMT uses more 'smarts' to tailor its signal to the channel.

The most significant sources of RFI are radio stations, because the ADSL band of 1 MHz sits on top of a portion of the AM band, and signals leak into the phone wires. These are extremely predicable (otherwise it would be hard to use your car radio), so the DMT modem puts signal power where it pays off most-at the receiver-and not, for example, where it will get wiped out by a radio interferer. While CAP merely 'powers through' the RFI, DMT simply avoids it, putting energy into frequency areas that can use it instead of wasting energy fighting a powerful AM broadcaster.

Thus in the ADSL environment, DMT copes with RFI in a much more efficient and intelligent way than CAP. And it copes with impulse noise better than CAP or QAM, because it's inherently more immune to it.

The standards process works through consensus: experts from across an industry meet, debate, and share opinions and research in order to arrive at a solution that's greater than the sum of its parts. The disadvantage is that it can be time-consuming; a consensus-based solution will arrive on the market later than an in-house approach that doesn't concern itself with the needs of others (since CAP technology doesn't adhere to any standard, proprietary chipset solutions shipped before optimized standards-based DMT chipsets. However, without interoperability, the usefulness of these proprietary products ended when the trials ended. Now, standard DMT chips are available).

However, there are far-reaching benefits. When a standards-based solution does arrive, it is supported by a wide base of manufacturers, bringing interoperability between them, and fruitful competition among them. Further, the process of experts from a variety of companies working together and discussing the technology usually leads to a solution that's much better than any one of those companies could have developed alone (the classic example: V.34 modems).

Consequently, both ANSI (T1.413 for the U.S.) and ETSI (TR238 for Europe) adopted DMT and the same standard. This is now endorsed by the ITU: "Q4/15: The initial work will focus on developing Recommendations for ADSL and HDSL based on the existing ANSI Standard T1.413 [for ADSL] and ETSI ETR152 Edition 3 [for HDSL] respectively. These proposed Recommendations will adopt these existing standards by reference."

At least eight different companies have developed independent solutions, all to comply with the same definition. Issue 2 of the standard is now being prepared, resolving many minor editorial changes and updating the body to formalize new advances (e.g., protocols for rate adaptation, ATM cell transport or higher-speed ADSL) that were implicit in the previous draft, but may not have been codified. None of these changes hinder compatibility or interoperability.

In contrast, CAP remains a single-source proprietary technology that not has been standardized. Late last year, an ad hoc group was set up to document a single-carrier approach. To date, the group has made no significant progress. In fact, it wasn't until May 1997 that the most fundamental choice was made: because there was no agreement about whether CAP or QAM was the better choice, all modems will be designed as dual-mode to support both, inevitably requiring greater complexity.

Given the amount of other details still to be resolved (it took more than two years to document DMT), it is obvious that single-carrier solutions will not be codified in the near future. This virtually guarantees there will be no interoperability or backwards compatibility between them.

Unfortunately, in the two years that CAP chipsets have been shipping, interoperability has never been demonstrated-not even between two modem manufacturers using the same chipset! This is astonishing, and it's primarily because CAP is not a specified standard, so many essential features such as error-correction must be individually developed by each manufacturer rather than within the chipset (currently, CAP is a proprietary single-source technology, so interoperability between different chipsets is meaningless anyway). In contrast, interoperability between modems from different manufacturers based around the same technology has been demonstrated by DMT suppliers.

Because DMT is so adept at matching its transmission channels by varying each of the 200-plus tones independently, it is both more efficient in its use of bandwidth, and delivers higher performance under any realistic circumstance.

Since the Bellcore Olympics, CAP modems have avoided independent tests that produce open, public results. For example, GTE published an audited performance test, and Tele.com magazine performed a Consumer Reports-style trial of a number of modems from different manufacturers. Several DMT modems (and even one HDSL modem) were tested. All CAP suppliers declined.

Where test results have emerged, it is clear that CAP consistently under-performs and is less robust. For example, Network Computing recently completed a trial (testing was performed by the independent MCI Test Labs):

Since some don't agree that DMT is better than CAP, we decided to include both in our tests. Although the modem units used in our tests are early releases, all performed at exceptional levels. Overall, we found that the DMT-based ADSL modems were more robust in signaling and were able to perform over longer distances (up to 18,000 feet).

In fact, we found that most of the modems achieved distances of up to 15,000 feet (-26 dB). The DMT-based ADSL modems in our tests were able to operate rate adaptively up to our maximum cable distance of 18,000 feet (with a measured line attenuation of 31 dB).

The CAP-based modems operated at a full speed of up to 4Mbps downstream and 422Kbps upstream until 12,000 feet was reached for one of the modems, and 15,000 feet for the other modem, which reached lower speeds of up to 2.2 Mbps.

[The DMT-based modems had maximum speeds of up to 8Mbps/768Kbps and achieved higher speeds than CAP at comparable rates]

In other words, despite the alleged maturity of CAP, an impartial public test found DMT modems faster, more robust and considerably longer in reach. While CAP modems needed different models to be optimized for reach or rate, the DMT modems (from different manufacturers) rate adapted well, achieving either 100% faster rates with at least 50% extra reach than one specific CAP model, or four times the maximum speed with 20% extra reach.

The maximum reach was limited by the test environment; however, the DMT modems would have gone further-they still had plenty of scope for rate adaption and further reach beyond 18,000 feet. This is entirely consistent with predictions.

However, this doesn't mean that at the same data rate, under the same test conditions, CAP will use less power. In fact, a fair comparison will show that DMT ADSL actually requires about the same power as CAP—or less. Here's why:

The signal processing complexity of the two is comparable in terms of MIPS or die area, although DMT is more 'digital oriented' than CAP (which uses more precise analog filtering), and will benefit faster from process technology advances.

The energy used in the driver dominates the system (roughly 50% of the total power used by the chipset. This depends on the power spectral density (how many watts are put into each Hz of bandwidth) and the bandwidth used. Since CAP is less well matched to the line, and may not benefit from FEC, to obtain equivalent data rates it needs to relax the psd mask and transmit at a greater power spectral density. In effect, it squeezes more energy into each Hz to overcome noise and the weakness of its implementation, at the expense of power dissipation and cross talk.

It is hard to generalize, as each system can set its own driver power to meet its specific needs. However, according to the specifications in the ad hoc proposal a CAP solution transmits at a much higher power level per Hz. The CAP signal at up to -34dBm/Hz, where a DMT signal is limited to the lower -40dBm/Hz. CAP's extra 6dBm/Hz means more signal power must be placed per Hertz, all the time.

What about the peak-to-average ratio (PAR) or crest factor? The power discussed above is the average level drawn continuously by the driver and sent on to the line. At times there will be a need for more, as the signal hits a particular pattern, and the peak power will be very much higher than the average level. For CAP, the ratio between these two is 4.0 (in other words, the peaks are four times the level of the average); for DMT the ratio is 5.3. In a piece of marketing spin it is sometimes claimed that because DMT has a higher crest factor it draws more power. This utterly ignores the fact that PAR is a ratio and that without knowing the average, you can make no conclusion about the result.

The actual specifics, for two systems with comparable net (after overhead) throughput are:

Pseudo-scientific references to 'peak to average ratio' notwithstanding, the simple physical fact is due to its lower efficiency, for like data rates and equivalent loops, a CAP system will require more power than a DMT system.

While most DMT systems are designed for maximum rate and reach, per T1.413, it is possible to dramatically reduce driver power for other configurations, if rate or reach can be relaxed. In these cases, the signal processing and error-correction "smarts" are employed to reduce power, rather than squeeze out the last drop of performance. For example, reducing the data rate to achieve 1.5 Mbps at 12,000ft could save more than 1W in the driver.

Additionally, 5.3 is PAR for a raw or 'naïve' DMT solution. A number of digital techniques exist for gain scaling and peak mitigation that can significantly reduce this, which would reduce the peak power needs. Once more, DMT uses digital technology and algorithms to significantly improve system performance over the 'obvious' implementation. Other technologies don't have the scope to use these opportunities, and faultfinding may arise from critics' lack of familiarity with such techniques.

This psd mask was defined by T1E1 for ADSL as a generic technology not specific to any one line code, in other words, whatever the implementation, it must meet the requirements and strictures to not cause interference. Unfortunately, the current CAP document and existing implementations do not follow this mask. Instead of stopping at 138 KHz, the upstream continues up to 180 KHz. This will sabotage the performance of standard DMT to such an extent it is unlikely it could be deployed in the same binder for normal length loops. A standards-compliant DMT system is better behaved, and will not interfere with other technologies, including CAP systems.

Similarly, on the downstream, CAP continues out to 1.5 MHz, instead of 1.1 MHz. This extra bandwidth places it directly into the VDSL band, causing significant interference and cross-talk there. It appears that a single ADSL system with these properties would dominate all other noise sources and potentially make VDSL unusable on all but the very shortest of loops.

In addition to choosing a robust modulation technique, error-correction coding provides a powerful technique to protect against noise and interference. Virtually every communication system in the world today employs error-correction and coding, and it is almost inconceivable that any modern system would ignore the benefits of including them. Probably the only exception is CAP chipsets: the existing chipset supports error-correction only on the downstream (in other words, the upstream is completely unprotected). Worse, the proposed ad hoc report drops even that, so there's no standard defined error-correction at all.

Since error-correction is essential in any practical ADSL system, the fact that it is not implemented in the CAP chipset has two critical implications:

1. It must be implemented in the modem outside the datapump chipset. This complicates the system design significantly. It is also wasteful, expensive and inelegant, especially since error-correction can be efficiently implemented in a comparatively small silicon area as part of a datapump.
   
   
2. It makes interoperability virtually impossible. Error-correction is so fundamental to a connection that if it's removed from a "standard" and made part of each modem manufacturer's individual system, it is equivalent to abandoning interoperability on anything other than a trivial system (one with no error-correction is not realistically deployable).

It is important to remember this when comparing data rates. The 6.1Mbps defined in ANSI T1.413 is net effective payload—the actual line rate, with overhead, is higher (approximately 6.9Mbps). In contrast, the "7 Mbps" of CAP is a gross rate; the effective data rate will be much less, since all error-correction will have to be subtracted from this. However, since error-correction is not standardized, it is difficult to give a specific answer as to what this will be, and the result will vary with the implementation.

DMT achieves rate adaptation easily and flexibly. It uses hundreds of degrees of freedom (subchannels) to accomplish it, and delivers the maximum data for any given line. This allows the support of higher rates over shorter loops (>8 Mbps), or sub-rate connections at very long reach (perhaps a few hundred Kbps over many miles, load coils permitting of course). CAP can support rate adaptation by only varying the constellation and the bandwidth of a single carrier. This requires very careful analog design, and the rates have much poorer granularity. DMT steps smoothly in 32 Kbps steps from 64 Kbps to >8 Mbps actual payload, while CAP has coarse and erratic steps from 640 Kbps to 7 Mbps (gross rate without coding).

DMT is like a mountain bike: it has lots of gears to adapt to different terrain. Some gears are suited for low rates and long reach of rural areas, others for very high speed (T1.413 Issue 2 allows up to 16 Mbps) on the short loops of urban environments. This is especially important when discussing rate adaptation for lower rates and longer reaches, a key concern for U.S. operators. While a 'coarse' step of >300 Kps may not matter in stepping between multi-megabit speeds, it definitely matters at lower rates. In fact, CAP simply can't support these applications: its rate adaption steps go from 960 to 680 to 640 Kbps, then drop off to nothing-meaning that no service at all can be delivered!

This is disastrous for operators who need to cover long reaches and rural areas, even if at lower rates. In cases involving distances of several miles, a rate-adaptive modem that can support a few hundred Kbps is a major attraction. CAP cannot serve these remote customers, whereas DMT will steadily scale to support very long reach, its rate dropping smoothly as reach is increased, from 960 Kbps to 928, 896, and so on down to 64 Kbps at very long range, allowing an operator to cover more area and offer service to more people.

This becomes even more important when you consider the effect of area. Suppose that, due to DMT's greater efficiency and channel matching, it can deliver a like rate at 25% greater range than CAP. This allows it to serve almost 60% more area! When you consider the many thousands of feet that DMT can serve at low-speed/long reach, that's a huge area—and a huge number of subscribers—that can't be served by CAP.

One final point: DMT varies the rate by digitally adjusting the individual sub-carriers. It is very flexible and efficient, and easily scales (costs reduce) with process technology and Moore's law. In contrast, CAP uses an analog technique (with different filters) that is far less flexible, and less amenable to cost reduction.

CAP systems were originally built for video only as evidenced by their minimal upstream capability (16 Kbps, increased to 64 Kbps in 1996—enough to pick a movie but not enough to provide PPP handshakes during an IP session). Not until two years after release of the first products was the upstream redefined to support data services. The original DMT standard, as defined, specifically recognized data services (Annex G of T1.413-95 explicitly discusses data access, remote LAN access and telecommuting as key applications, and their requirements), and defined an upstream ratio accordingly.

DMT has higher bandwidth efficiency, which translates to higher speed (and bandwidth is like PC memory: we always seem to need more) only at the expense of end-to-end latency, which doesn't matter for internet access. Within the standard document, the system is defined to provide a 10:1 ratio in data rates, which studies have confirmed is optimal for internet access.

(CAP is inherently suitable for internet access? Its supporters spent much of 1995 and 1996 struggling to support internet services with the antiquated upstream of 64 Kbps, while trying to justify a 20:1 or even 30:1 ratio as acceptable.)

Further, the latest version of the DMT standard goes even further, explicitly documenting the ways that DMT and T1.413 will grow, while remaining backward-compatible, to support rate adaption, ATM and packet mode (IP and frame relay) data services in an efficient, versatile and interoperable way. These developments have no counterpart within the still-incomplete CAP specification.

It is claimed that DMT is heavily patented, and enclosed in a thicket of intellectual property. In fact, the situation of both DMT and CAP is remarkably similar. The principle difference is that DMT is documented in a standard, open forum, and independent vendors can use this to design compatible systems. However, the fact that all the details for DMT are publicly available encourages independent developers. Several developers (including Alcatel, ADI/Aware, and Orckit & Pairgain) have, completely independently, developed systems designed to the T1.413 specification.

In contrast, CAP has been a proprietary, 'closed' architecture, confirmed by the fact that not a single alternative supplier has delivered an open-market alternative.

DMT has long latency (2 ms). If this is important, another technology is more relevant (e.g., HDSL2 uses single carrier technology). However, for ADSL and its intended applications-internet access and broadband services-latency is a non-issue.

One of the strengths of DMT is its versatility, rate adaptability, and flexibility in coping with a vast range of environments found in the 700 million existing copper lines (efficiently matching its properties to the channel, and coping elegantly with noise). In some cases, where that versatility is not required, a less adaptable technique may be more appropriate.

Finally, in applications where extreme power efficiency is crucial and takes precedence over data-rate or information density (e.g., battery powered wireless with data rates of just a few Kbps), single carrier techniques can be optimized (e.g., OPQSK or GMSK used in digital cellular) for efficiency. However, the high data rates required of ADSL make it an entirely different application (in any case, the CAP used in ADSL is a different technology; it cannot achieve this optimization).

It is notable that most of the suppliers of DMT, including ADI, Alcatel, Motorola, and Orckit & Pairgain, are not confined to a single choice: they have the experience and skills to manufacture a variety of products using a number of different line codes. This suggests they are well positioned to recognize the best technology for an application. And each of them selected DMT.

We believe the standards process and industry experts have done a good job over the past five years of discussion. They concluded that DMT is the optimum technology for ADSL. Why? Because in the critical areas-communications speed, bandwidth efficiency, spectral compatibility, performance, robustness and power consumption, and delivery of a better service to more customers-DMT has proven superior to alternative single-carrier technologies.