November 6, 1997

Modem technologies: the choice is yours

Maury Wright, Technical Editor

The 56-kbps wars may garner the headlines, but the real excitement in modems lies in the range of technologies you can use to implement them. Chip sets are still the first choice, but burgeoning host-based implementations are closing in quickly. Designers should also start thinking about a merged modem/audio subsystem.

Modems are standard in almost every desktop computer and can be invaluable in embedded applications. Systems ranging from smart soda machines to remote data-acquisition systems all make good use of modems. You might think that such widespread usage would lead to a relatively standard modem implementation. Actually, however, system designers must choose from widely differing technologies--fixed-function chip sets to host-based signal processing (HSP)--to find the best match for an application. Furthermore, in the never-ending pursuit of lower prices, designers may need to consider emerging technologies that integrate audio and communication functions into one subsystem. They also have to sort through a virtually irrelevant propaganda war over emerging 56-kbps-modem technologies (see box "The ultimate red herring: 56-kbps modems").

Evaluating complex signal-processing-based technologies, such as those a modem requires, is never simple, but the modem technologies are even tougher to evaluate because of the different schemes available to designers. Before you can even think about typical criteria, such as price, performance, and reliability, you must fully understand all the implementation options. Your choices fall into the following sometimes-overlapping categories (see Table 1 for vendors):

• Fixed-function chip sets that have a DSP-based data-pump IC that executes the modem algorithms and a µC that handles control functions. These functions include the industry-standard Hayes AT command set, error correction, and data compression;

• Controllerless implementations that have a DSP-based data-pump IC for the modem algorithms yet rely on the host CPU for control functions;

• General-purpose DSPs that can execute the modem algorithms and, in some cases, the controller functions as well. Optionally, you can use a µC or the host to handle control functions; and

• HSP modems that execute both the modem algorithms and the control functions on the host CPU, relying on little more than a line codec, a bus interface, and some buffers in hardware.

The traditional fixed-function chip set commands the greatest market share. It may be the safest choice when you must have a modem up and working quickly, and it is often the only choice for an external modem designed to work with any type of computer. Rockwell reigns as the overall market leader in modem chips because of its market dominance for fixed-function chip sets, and Lucent Technologies and Cirrus Logic join Rockwell in offering such chip sets. Texas Instruments also competes in this market, because its modem product is based on a C5x DSP and can include the controller function executing on the DSP. Motorola has also made a lukewarm entrance into the fixed-function-modem business using its 56303 DSP, but the company has not yet announced a design win.

Complete modem chip sets offer a safe design path because you have little designing to do. Vendors of these chip sets typically offer detailed reference designs that a customer can use to begin manufacturing. These designs assure the customer that the end product will meet FCC or other regulatory standards. Moreover, box-level modems that use complete chip sets don't rely on any specific operating system or host processor; you can adapt them to almost any computing environment.

Despite a long history, complete chip sets may soon stop dominating the modem market for several reasons. First, the computing world is increasingly migrating to a Windows/Intel (Santa Clara, CA) environment. The "Wintel" platform is one of the most popular choices among embedded-system designers today. Once you identify Wintel as your target, you can reliably move toward a controllerless or HSP-based design and derive significant benefits. Second, the demand for external box-level modems is shrinking, and the demand for retail after-market modems in general is shrinking. PC OEMs today invariably include an internal modem in all systems, so only users needing to upgrade to a faster product buy retail modems.

Lose the controller

The trend of PC OEMs' directly integrating modems has led the move toward controllerless designs. Controllerless designs emerged around 4 years ago, but the technology created some problems. First, the control-code, host-controller concept was new, and it took awhile for controllerless pioneer Lucent and others to work out the kinks. Second, the technology sold through the retail channel, which was then the dominant channel for modems. Unfortunately, retail buyers had a hodgepodge of OS versions and CPU--some of which weren't powerful enough to host the control functions.

PC OEMs now have plenty of CPU horsepower for modem control, and they love to eliminate the cost of a µC. They also control the version of the OS that ships with each system and can therefore eliminate driver problems. Moreover, controllerless modems have improved and even sell successfully through retailers. Controllerless-market leader Lucent Technologies, however, recommends the technology mainly for system manufacturers.

You can also buy chip sets based on off-the-shelf DSPs, and you can use host-based control. Alternatively, with DSPs, you can use a dedicated µC for control, or you can implement control functions on the DSP itself. You might wonder why DSPs deserve a category independent of the other ICs, because the fixed-function and controllerless chip sets are essentially customized DSPs. The difference lies in the other features that typically do not exist on a DSP but that vendors have integrated onto modem chips. For example, along with a DSP core, modem chips typically include ISA- or PCI-bus interfaces, UARTs, or RAM, whereas DSPs require you to implement these functions in another IC. Only TI and Motorola actively market their DSPs for modem applications, but you could realistically use almost any DSP, including those from Analog Devices, or even DSP cores such as the Oak and Pine cores from the DSP Group. Independent software vendors such as Vocal Technologies and DSP Software Engineering offer modem algorithms for these DSPs.

PC OEMs adopt HSP

The hottest trend in modems for the next year, however, will likely be HSP modems. The PC OEMs love the elimination of the µC and are salivating over eliminating the DSP IC. PC-Tel claims to have shipped 2 million HSP modems, mainly in systems from less well-known PC vendors. This year, expect tier-one PC makers to adopt the technology in some systems. PC-Tel is the leader in HSP modems and is the only vendor to ship significant quantities. Several other vendors, including Motorola, ESS Technology, and SmartLink, are just starting to ship HSP modems.

An HSP modem typically includes one IC that includes a PCI- or ISA-bus interface and buffer memory that ensures relatively smooth data flow in the non-real-time Windows environment. The vendors supply the modem algorithm as a virtual device driver (VXD) that executes at Ring 0 of the Windows OS, thereby enabling the modem algorithm to preempt user tasks. The vendors all claim that HSP modems require approximately 25% of the CPU's cycles in a typical 200-MHz PC. That number drops to approximately 10% on a 266-MHz Pentium II, which will soon be a mainstream processor. Most observers agree that 10% is the magic number after which a user doesn't notice any impact on system performance due to modem overhead. In an ironic twist, however, PC OEMs would prefer to ship HSP modems in lower priced, more cost-sensitive PCs, which generally have slower µPs.

HSP modems aren't strictly a Wintel phenomenon either. For example, Advanced RISC Machines (ARM) offers HSP modems for its RISC µPs. Those ARM µPs target handheld computers, among other markets, and an HSP modem would fit well in the market for mostly single-tasking, price-sensitive handheld computers. If you use something other than Intel or ARM µPs, look for a modem algorithm from a third party before you dismiss the technology.

Once you understand your modem-technology options, you should turn your attention to the varying performance, price, and extensibility attributes of each of the technologies. Starting with performance, you might think that HSP modems perform significantly worse than DSP-based chip sets, but that's not necessarily so. Warren Henderson, president of modem-testing expert Henderson Communications Laboratories, just evaluated a PC-Tel modem and found that in data-pump throughput, the HSP modem compared favorably with DSP-based modems. Henderson reveals that HSP-modem throughput can suffer if the user runs multiple tasks while using the modem. The degradation results from an impact on the host-based data compression and error correction rather than from the modem algorithm. The control functions don't run at Ring 0 and, therefore, don't enjoy the preemption privileges available to the modem algorithm. In fact, controllerless modems can suffer similar performance impacts.

Vendors guard data

Modem performance is a sensitive subject for modem-technology vendors. The basic method of testing modem performance centers on the Telecommunications Systems Bulletin (TSB) 37A and 38 testing documents, which the Telecommunications Industry Association (TIA) publishes. The documents specify a range of modem tests that vendors must perform over a variety of simulated line conditions, which should reveal the throughput a modem can achieve. Most designers, however, find the TSB 37A and 38 tests beyond their abilities and budgets.

Modem-technology vendors internally perform these tests but rarely disclose results in public. Only Cirrus Logic willingly provides some results, and, not surprisingly, its chip set tested better than modems based on other chip sets. Modem-technology vendors claim that they don't publish such test results because the design techniques of their customers often determine how well a modem performs. A more robust pc-board design or higher quality components, such as capacitors, can improve performance. Moreover, larger modem manufacturers often add custom features in the control functions and occasionally customize the modem algorithms.

It seems logical that the modem-IC vendors could release test results of their reference designs as typical performance. The vendors, however, have other reasons for secrecy. Modem-IC vendors constantly tweak code to improve performance when TSB tests reveal that a competitive product performs better in some way. The constant tuning of code makes disclosure of testing results problematic. You can contract with Henderson Communications Laboratories to perform tests for you. In some cases, you can even buy copies of reports the company has prepared for other companies. For example, US Robotics commissioned Henderson to test US Robotics' TI-based modems against a number of competitors and authorized Henderson to sell the report for $12,000 to interested parties.

You can make a few generic observations about performance in the modem-technology categories. On a lightly loaded system, a controllerless or even an HSP modem can offer higher data rates than can a fixed-function chip set. This irony results from the fact that the fixed-function chip sets use a UART through which all host and modem data transfers must pass. The UART is a holdover from the original external modems and is necessary to provide the industry-standard register set for the Hayes AT command set. That serial port effectively limits throughput between the host and modem to approximately 230 kbps and, often, even slower rates. The controllerless and HSP modems have no such roadblock and can transfer data at PCI- or ISA-bus speeds. On the other hand, the performance of fixed-function chip sets never suffers because of other active tasks running under Windows.

Design for easy upgrades

In reality, only a few large OEMs make decisions based primarily on performance. More often, interrelated criteria, such as upgradability and price, are more important. In fact, the 56-kbps movement has solidified the only outright mandate in modem design: You must design for easy upgrades. Upgradability became an issue during the drawn-out development of the V.34 standard. Pre-V.34 proprietary schemes, such as Lucent's 19.2 -kbps V.terbo and Rockwell's 28.8-kbps V.fast, resulted in user demand for an upgradable modem. Many modem vendors had to promise a free V.34 factory upgrade just to lure customers to the new products. It didn't take the IC or modem manufacturers long to realize that they needed to offer user-upgradable products.

With no 56-kbps standard on the horizon, all the vendors selling modem chips have an upgradable plan. The way vendors achieve upgrades, however, affects the complexity of the design and the bill of materials required--both of which impact price. The fixed-function chip sets are at a disadvantage because they usually require flash memory to store code changes. Controllerless and HSP modems can store upgrades on a user's hard disk. In the controllerless case, you download the code changes to the modem IC. With fixed-function and controllerless modems, you should also find out whether additional RAM is necessary to store the upgraded code during execution.

A number of other factors affect the cost of a modem implementation. An average V.34 modem chip costs $20 to $30 in the 10,000-lot quantities typical of the PC market. Right now, you can expect to pay a $10 to $20 premium for x2 or K56flex support for 56-kbps operation. At the high end of those price ranges, you find modems with dedicated controllers or DSP-based modems that require external static RAM. At the low end, you find the controllerless offerings, and HSP modems should cost less than the stated range. PC-Tel quotes a price of $20 (10,000) for its HSP modem, including the bus-interface/buffer IC.

Unfortunately, modem pricing is not straightforward for anyone other than the major semiconductor players, including TI, Lucent, Rockwell, and Motorola. These semiconductor vendors all own patents that are crucial to V.34 and faster modems, and they all have cross-licensing agreements in place. These intervendor agreements allow the companies to sell modem ICs and indemnify their customers against any patent issues. Smaller companies that lack large patent portfolios in the modem area have to pay an outright license fee approaching $10 for each modem IC or HSP modem they sell. The license fees, however, are necessary for the smaller companies to offer indemnification. Of course, what goes around comes around, and the big guys may find they need licenses from the little guys down the road. For example, PC-Tel was just awarded a patent for HSP modems, and the IC vendors may find that they need to offer HSP modems in the future. In any case, make sure you receive indemnification and factor in the effect on price to ensure that you make the best choice before buying any modem technology.

Flexible design cuts cost

It's possible that a designer, especially one working for a system OEM, can't make a single modem-implementation choice that fits all applications. For example, a PC manufacturer may want to package a controllerless modem in some systems and an HSP-based modem in others. Moreover, the manufacturer may require ISA-based modems in some systems and PCI-based modems in others. In such cases, flexibility of implementation may become the most compelling attribute in your selection. Moreover, a flexible implementation can cut costs in the long term when you consider manufacturing and inventory costs.

Upstart modem manufacturer ESS Technology offers OEMs just such flexibility with a pair of pin-compatible ICs, both of which integrate PCI and ISA interfaces. The ES2820 integrates a DSP-based data pump and all the SRAM necessary to execute modem algorithms. The chip targets controllerless designs; can handle telephony functions, such as a full-duplex speaker- phone; and is upgradable to 56-kbps operation. Meanwhile, the ES2808 offers pin compatibility with the ES2820 yet omits the data pump. The ES2808 also includes the large buffers that are typically necessary for an HSP modem.

The ES2820 and ES2808 combination offers OEMs plenty of flexibility. For starters, designers can use one pc-board layout that allows the manufacturer to decide between an HSP implementation and a controllerless implementation as late as the board- assembly stage of manufacturing. The two chips can use identical codecs and can be equally effective when mounted on motherboard or add-in cards.

The ESS Technology combination also allows manufacturers to inventory just the two ICs yet support both ISA and PCI designs. The Intel-based PC and workstation world is quickly moving away from ISA. For example, the new TDZ-2000 Pentium-II Workstation from Intergraph (Huntsville, AL) offers a single slot that can be ISA or PCI; the remaining six slots are PCI slots. Lower end systems will certainly include ISA slots for some time, but it's clear that ISA will slowly disappear from systems. With the ES2820 and ES2808, manufacturers can support both types of modems and still be prepared for the transition to PCI.

Ultimately, however the most successful modem-technology vendors will be those that can directly lower implementation cost and still maintain functionality. Vendors take two approaches to lowering modem cost. First, some vendors plan to develop a true single-chip modem, thereby taking the traditional approach to lowering cost by integrating more into an IC. Other vendors plan to integrate different functions, such as modem and audio, into one subsystem to lower cost. One other approach to integration involves supporting higher speed communication schemes as well as modem schemes (see box "ICs bridge the gap to ISDN and ADSL"). You should consider all of these approaches for any new designs because these technologies will ship late this year.

Unlike earlier attempts to meld audio and graphics subsystems, audio and modem subsystems make a good match for combination. Both require work in the relatively low-frequency ranges, similar data rates, and similar signal-processing resources, and both must connect to peripherals, such as speakers. Intel also encourages this trend with its Audio Codec 1998 (AC '98) specification, which mandates one chip capable of performing both the audio- and the communication-codec roles. Such codecs will appear early next year. Meanwhile, the traditional audio- and modem-technology vendors must develop one DSP-based resource that can execute both audio and modem algorithms. The vendors must also develop a scheme to offload appropriate functions to the host and perhaps develop some type of dynamic load-balancing software for the DSP.

The move toward a merged audio/ modem subsystem could significantly change the modem-vendor pecking order. Rockwell has had a well-established audio group for several years and sells some audio-codec products, but it hasn't publicly offered its audio/ modem road map. Cirrus Logic has its CrystalClear (formerly, Crystal Semiconductor) audio technology but no public audio/modem road map. Lucent, meanwhile, has no real presence in audio technologies.

Surprisingly, however, you can already buy some audio/modem combinations. Perhaps ESS Technology will emerge as a leader in the audio/modem market, because the company has an established position with PC OEMs with its audio products and now offers perhaps the widest choice of modem technologies of any company. ESS was the first to introduce an audio IC designed for a combined modem and audio subsystem. The Maestro-2M, which ESS introduced this fall, includes the latest in audio technology, including 3-D positional audio and a 64-voice wavetable synthesizer. The IC also includes a port that connects directly to the company's 2820 or 2808 modem ICs so that you can easily integrate controllerless or HSP-based modem functions. The Maestro-2M costs $22 (10,000). ESS Technology also indicates that it will move both audio and modem algorithms onto one DSP in its Maestro-3M IC and hopes to ship samples by the end of the year.

Modem shares audio codec

Not all modem-technology vendors believe that you need a specialized audio IC to integrate an HSP-based modem with the audio subsystem. Start-up SmartLink ships Modio, a software product that doesn't require the typical bus-interface and buffer IC that HSP-based modems use. Instead, SmartLink relies on the audio-codec IC that's typically included in a system for the interface and buffering. The company's Modio approach requires an off-the-shelf multiplexer IC that costs less than $1 to connect to the output of the off-the-shelf audio codec. The Modio software and the multiplexer IC allow simultaneous audio and modem operation through the single codec. The company admits, however, that audio quality degrades during a modem connection. Still, the software may offer the most cost-effective way to add modem functionality to any system requiring audio support.

To date, SmartLink has verified that the Modio software works with the Analog Devices AD1821 codec and the Cirrus Logic CrystalClear Audio 4236, 4237, and 4238 codecs. To handle intellectual-property indemnification, SmartLink plans to license the Modio software to codec vendors that have cross-licensing agreements in place. Analog Devices has contracted with SmartLink to offer Modio as an option to AD1821 customers and calls the $17 (10,000) combination SoundComm. Shark Multimedia recently became the first company to announce a board-level, Modio-based product that it hopes to sell to system integrators and OEMs. Called the Basking Combo, the product supports 16-bit sound and modem operations, including 56-kbps support. Shark Multimedia plans to offer the product to OEMs for $40 to $50 based on volume.

The first single-chip modem

Surely audio/modem combinations will prove popular, but other companies will undoubtedly pursue a true single-chip modem. Expect Analog Devices to be first to market with such a product. The company is in preliminary discussions to produce an IC that integrates a PCI or an ISA interface; a DSP; all required memory; and line, telephone-handset, and speakerphone codecs. The IC would rely on the host for control functions and require nothing but a standard data-access arrangement for a modem implementation.

Analog Devices is developing the IC in partnership with a modem vendor, and that vendor will have exclusive rights to the IC for six months. Moreover, the modem vendor is developing the key algorithms and may not license them to other modem designers. Still, the IC essentially relies on an AD2181 core, so designers familiar with that DSP could roll their own code. Moreover, companies such as Vocal Technologies offer modem algorithms for the 2181, so the IC will be a viable choice for new modem designs starting early next year.

The single-chip modem will bring an abundance of economies to modem designs. For instance, Analog Devices believes that the cost premium for the new IC will be less than $10 when compared with a codec such as the AD1821. For high-volume manufacturers, the new IC should result in a bill of materials that is less than $20 for a DSP-based modem. The single-chip design will also allow designers to transition from typical six-layer modem pc-boards to two-layer boards, move to a single-clock design, save pc-board real estate, and use 50% less power.

| EDN Access | Feedback | Table of Contents |