10-Gbit/sec modules: after when, comes three more Ws
Here are some of the different flavors of 10-Gbit/sec fiber-optic modules, along with the whens, whys, whos, wheres, and hows.
AARON SCHULTZ, Quake Technologies Inc.
Walking "the floor" at the 2001 OFC conference in Anaheim, CA, in March, it was impossible to escape the conclusion that widespread proliferation of data-communications 10-Gbit/sec optical links has begun. Through the myriad sights and sounds that would almost certainly inundate a newcomer to the fiber-optic communications industry, it was clear that many vendors are on the brink of having fully production-worthy, next-generation 10-Gbit/sec fiber-optic modules.
Even with the economic slowdown, data traffic still increases on the Internet. The need for an inexpensive expansion of the Internet infrastructure pushes the development of these next-generation fiber-optic modules. Tens of module companies have joined the foray of already-established module suppliers, as everyone gears up for a taste of the 10-Gbit economic potential.
Certainly, 10-Gbit/sec modules have been around for a few years. The difference now is that the development and deployment of less expensive, more commercialized 10-Gbit/sec modules for more applications is no longer in question. The questions are, what market forces stand to define the modules, what optical configurations will link the modules, and what technological challenges will influence design of the modules?
Looking through history, increased resources seem typically to beget the need for yet more increases in resources. The development of the fiber-optic communications business is not a special case. Companies and individual home broadband users have successfully latched onto the larger number of bits per second that are available, thanks to the 1- and 2.5-Gbit/sec optical links that have thus far been deployed from the previous generation of module development.
DSL, cable modem, and wireless modem networks rely in some way on a fiber-optic backbone, whereby aggregations of single digits of megabits per second amass at central communication distribution points. Fiber-optic links at 1 Gbit/sec or faster are already needed to support efficient rerouting of these lower-speed communication tributaries. And as 1-Gbit/sec optical links provide connectivity even within LANs, there is already a need for 10-Gbit/sec or faster uplink speeds to communicate from one LAN to another over a WAN.
Compact modules that carry more data traffic in the same or smaller physical space as the previous generation will be part of larger routers and switches to handle the uplinks. While several companies are working on all-optical networks, it's the 10-Gbit/sec fiber modules that convert between optical and electrical signaling that will pervade the market.
Who, then, provides structure and organization to the module business? Several agencies. For one, some 10-Gbit/sec SONET networks for long-reach (20-40-km) telecommunications already exist. To this end, specifications, albeit difficult to meet in terms of jitter, already apply as a precedent (e.g., Bellcore GR-1377-Core from Telcordia Technologies). This standard was used primarily for the development of more expensive modules that could affect the longer-range links.
At the same time, the Optical Interface Forum (OIF) seeks to study the interconnection between data networks and optical networks. Their proposals, such as the OC-192 (10-Gbit/sec) serializer/deserializer (SerDes) framer interface (SFI-4) and the four different very-short-reach (VSR) optical interconnection types (100 to 300 m), are helping the industry converge upon technically and economically feasible optical connects. As a direct result of their work, for example, several companies have promoted 300- and 200-pin multisource agreements whereby an OC-192 fiber module footprint is defined.
In the meantime, the IEEE is readying its 10-Gbit/sec Ethernet standard, 802.3ae. This standard attempts, according to the charter IEEE set forward, to achieve board market potential, compatibility with IEEE 802.3 (previous Ethernet), uniqueness to optimize for current-day market and technological issues, technical feasibility, and economic feasibility. In the past, fiber-optic modules based on optical and elec-trical specifications from IEEE were optimized for low-cost LAN applications. The 802.ae 10-Gbit/sec standard fuses LAN and WAN, as well as MAN, considerations to enable Ethernet protocol to penetrate more into the total 10-Gbit/sec distribution.
Additionally, the Fibre Channel organization, which previously developed a 1.0625-Gbit/sec storage-area network (SAN) protocol, will follow the Ethernet development by a six-month lag schedule with a new 10-Gbit/sec SAN proposal. Further 10-Gbit/sec developments from other agencies include parallel interfacing for backplane connections and other proprietary machine-to-machine interconnection schemes. In every application created by the implementation of one of these efforts, some kind of 10-Gbit/sec fiber-optic module will need to be produced.
Optical links can be divided into distance categories. Searching through OIF articles and IEEE 802.3ae work, there are similar classifications of links. As an example, consider a proposal from the IEEE.1 The Table (on page 202) from this proposal groups 10-Gbit/sec links by distance and suggests what types of physical media devices (PMDs) will be most suitable. Further descriptions of modules that comprise these PMDs follow.
While the Table may not represent exactly what types of modules fit into which link distances in the end, it suggests that different types of 10-Gbit/sec modules will serve different link implementations. Terms like "short-reach" (SR) and "long-haul" have different meanings, depending on which standard is being referenced (e.g., IEEE 802.3ae versus Bellcore GR-1377-Core). Correspondingly, different 10-Gbit/sec modules will appropriately fit different reach lengths, depending on markets.
People and companies are thirsty for 10 Gbits/sec of streaming data. Standards organizations are preparing specifications for 10-Gbit/sec telecom systems, data networks, SANs, and other interconnects. What then will 10-Gbit/sec fiber-optic modules look like?
One way to categorize these modules is by the optical interface. Specifically, how many wavelength carriers are needed to yield 10 Gbits/sec? For systems with more than one carrier, how many fibers, or how many wavelengths, and at what data rate per channel, are needed to equal 10 Gbits/sec? The many flavors of modules can be categorized into three types: serial, parallel, and WDM.
Serial modules. Figure 1 shows a diagram of a duplex fiber-optic module. The transmitter (Tx) integrated circuit (IC) and receiver (Rx) IC, and the communications lines that mate to the external system board through the connector, will accomplish whatever the specified functionality is. Some fiber module Tx ICs will simply drive a laser based on incoming serial data. Some fiber module Rx ICs will simply digitize the analog signal coming from the transimpedance amplifier.
In other cases, the Tx and Rx ICs will also perform serialization and deserialization, respectively (converting one parallel bit data stream at one speed to a serial bitstream at higher speed, or visa versa), or other more advanced functions. Inside the module, a transmitter optical subassembly (TOSA) and receiver optical subassembly (ROSA) contain optical devices (laser and PIN photodetectors) and focusing mechanisms.
The device in Figure 1 is a serial optical module, whereby all 10 Gbits/sec out of the module is carried by one wavelength into a single optical fiber, and all 10 Gbits/sec into the module is carried by one wavelength out from a single optical fiber. This serial optical transport is the most generally cost-effective method of transporting 10 Gbits/sec-or any data rate-of optical data in a link.
In previous fiber-optic-module generations, a single bitstream in a single fiber-optic cable conducted the full bandwidth of data. At 10 Gbits/sec, two issues may hinder the initial deployment of mass-produced serial modules: When will 1,300-nm, or even 850-nm, vertical-cavity surface-emitting lasers (VCSELs) be production-worthy? And what is the optical bandwidth of already-installed fiber-optic cable?
VCSELs have provided much excitement in the fiber-optic module-making business over the last five or more years. A VCSEL is a semiconductor laser diode that emits light in a cylindrical beam vertically from the surface of a fabricated wafer. This device can be manufactured using typical IC fabrication steps and tested on a wafer due to the upwards-pointing light cone emission. Hence, manufacture of VCSELs is generally less expensive than distributed-feedback (DFB) lasers. High speeds (>10 Gbits/sec) with small driver currents (<10-mA amplitude, <5-mA threshold) make VCSELs very attractive from a system performance standpoint.
The "holy grail" of OC-192 is the 1,300-nm VCSEL. Typical fiber-optic cables have a null in their loss characteristics at 1,300 nm. As a result, 1,300-nm optical links typically can be longer than shorter-wavelength links. The "earthly grail" (more likely to succeed earlier on) is the 850-nm VCSEL at 10 Gbits/sec. Already SONET systems use expensive DFB singlemode lasers, and there are hopes of using 1,550-nm VCSELs that benefit in link distance from erbium-doped fiber amplification at that wavelength.2
But for 10-Gbit/sec modules to be rampant (not just SONET very-long-haul links), there needs to be a production-worthy supply of low-cost singlemode or multimode lasers. If these lasers existed today, there would be no question that the optical interface would be serial. But since these devices may be difficult to obtain during the initial proliferation of 10-Gbit/sec modules, other optical interfaces (such as parallel and WDM) may serve to achieve the data rate in the interim.
The other issue with 10-Gbit/sec serial optical transmission is that the optical bandwidth of some of the already-installed fiber-optic cables starts to degrade-significantly enough that 10-Gbit/sec links may be limited to 100 m or even less with 850-nm lasers. In some cases, higher-quality cables must be added for new 10-Gbit/sec communications networks. For example, the High Speed Study Group of the IEEE is targeting for 10-Gbit/sec Ethernet over a 100-m multimode link distance with already-installed multimode fiber. This same fiber could carry 1-Gbit/sec multimode optical signals 300 or more meters, depending on the fiber's core diameter.
In any case, for VSR, SR, and long-reach, serial communication modules and serial optical links should, in theory, be the lowest-cost 10-Gbit/sec (or any Gbit/sec) solution, all else being equal.
Parallel modules. While the production issues for less expensive 10-Gbit/sec lasers suited for LAN, MAN, and SAN applications settle, parallel optical topologies have been developed for VSR applications. Parallel optics-physically separate fiber cables bundled into one aggregate transport-is a popular concept in backplane connections, where installing short lengths of specialized parallel optical cables is feasible. Conceivably, some LANs may be installed with parallel fiber bases, but initial development of parallel fiber has centered on backplane interconnects.
Several parallel optical modules are on the market. The 30-Gbit/sec fiber-optic modules use 12 2.5-Gbit/sec optical channels. The 10-Gbit/sec is achieved by using four 2.5-Gbit/sec optical channels. Inside each module, electromechanical schemes couple light from an array of lasers into an array of optical fibers and focus light from an array of optical fibers onto an array of PIN photodetectors.
Two of the challenges in developing parallel optical modules are how to yield production laser arrays and couple light from device arrays into fiber arrays within a small space. The module in Figure 1 is expanded such that there are N lasers, N photodetectors, and enough fiber cores. ICs with enough integration can handle drive of all lasers or reception of all signals. Figure 2 shows a possible configuration of components at the transmitter front end of a parallel fiber-optic module.
WDM modules. An alternative parallel optical scheme that can use already-existing LAN fibers is WDM (see Figure 3). Some companies have developed arrays of optical transmitters and receivers with many wavelengths-each slightly different from the others-and incorporated them into optical modules. Optical packaging concentrates the light beams into or out of one fiber.
This scheme has two advantages. First, like parallel optical modules, it solves the problem of unavailable production 10-Gbit/sec VCSELs. Second, since some already-installed multi-mode fiber de grades the bandwidth of 10 Gbits/sec even after 100 m, it would be advantageous to combine different 1.25- or 2.5-Gbit/sec optical signals that can travel 300 m or more, especially around 850 nm.
Several challenges have helped to determine the components of 10-Gbit/sec fiber-optic modules. In interfacing the fiber-optic module to a motherboard, the data-rate signals in previous-generation modules could be brought out to pins on a connector. Even at 1 Gbit/sec, two metal pins could reasonably mate a differential signal to a printed circuit board (PCB) and carry signals without too much signal degradation.
If there were reflections due to an impedance discontinuity, the time during which the reflections settle out would typically be much less than one bit-period (15-cm wavelength with dielectric constant ε (of 4 at 1 GHz is typically much longer than signal traces in a module). Additionally, loss tangents of PCBs at 1 GHz are such that signal bandwidth roll-off and amplitude loss are acceptable, even with the use of common, standard, inexpensive FR4 PCB material.
With 10 Gbits/sec and faster, the relevant frequency issues become more difficult. The 10-Gbit/sec signals are 5-GHz square waves. While no chips output perfect square waves, 10-Gbit/ sec physical-layer (PHY)-chip outputs are square enough to contain energy in several harmonics above 5 GHz. To avoid bandwidth roll-off, all signal conduits need to be able to carry at least a few of these harmonics.
Low-cost, highly manufacturable materials such as FR-4 have significant loss tangents at 5 GHz and beyond. At 10 Gbits/sec, unlike with past slower data rates, these bandwidth loss mechanisms are strong enough to degrade the signal significantly with even very short (1-inch) electrical paths. In some cases, materials such as N4003, Duroid, or Teflon are needed to ensure adequate high-frequency PCB performance. Any kind of impedance mismatches (e.g., pins carrying 10-Gbit/sec electrical signals) at 5 GHz and above give rise to waveform degradation that closes an eye-diagram.
Many 10-Gbit/sec modules will likely not use such metal-carrying 10-Gbit/sec data to interface to the system boards. Instead, many modules will use a solder ball grid to connect to systems boards. Since systems boards cannot carry 10-Gbit/sec data more than about 1 inch, many 10-Gbit/sec modules will connect by way of a lower-speed parallel data stream. For example, some OC-192 modules use the OIF SFI-4 16x622-Mbit/sec interface. Modules such as these require PHY chips to convert the high-speed data stream to a lower-speed parallel stream, and visa versa. Gone are the simple form factors of the 1-Gbit/sec modules! The 10-Gbit/sec modules cannot avoid added input/output complexity, since the only feasible place for 10-Gbit/sec electrical signals is on a well-designed PCB inside the optical module.
Custom PHY ICs for fiber-optic modules have historically been implemented in silicon bipolar technology. Exactly how advanced a process was needed depended on power, speed, and integration requirements. BiCMOS processes were sometimes used for higher levels of integration or circuit improvements. In cases that required frequency of operation beyond the capabilities of bipolar technology, III-V devices such as gallium arsenide (GaAs) were used. Silicon germanium (SiGe), a high-speed process with low power and an available BiCMOS mix, was another possible alternative where silicon technology did not perform adequately. Only recently have CMOS process linewidths become small enough that the rise/fall time and jitter performance can be used in some PHY IC applications where low power versus performance is the tradeoff.
The question, then, is which technology will be used in the 10-Gbit/sec PHY chips? As the power usage and cost of some SiGe solutions are reasonable, and as the mixed-mode technology affords high levels of digital and analog function integration, it is likely SiGe will, perhaps by necessity, be the technology of choice, at least in the first batches of 10-Gbit/sec optical modules.
There is some doubt that CMOS edge speeds can be fast enough and whether CMOS jitter performance will meet SONET specs, although there are efforts to develop 10-Gbit/sec CMOS chipsets. Other solutions using GaAs and indium phosphide are being considered for their speed, especially as the industry looks forward to 40- and 100-Gbit/sec serial data rates. Whichever technology is used, the 10-Gbit/sec fiber-optic module will almost surely feature a technological shift from previous generations that could rely on silicon bipolar processes.
Many module makers faced the electromagnetic interference battles at 1 Gbit/sec. From this experience, 10-Gbit/sec modules will likely be designed so in an application, radiated emissions will not be a surprise stumbling block at the end of a project development cycle.
A new hot technical issue to 10 Gbits/sec is heat dissipation. Integrating more functionality at higher speeds into smaller package sizes has its price-more heat per volume generated. Some initial SerDes IC chipset introductions, for example, generate enough heat to warrant a heat-sink. While decreasing supply voltages and design re-spins will help reduce power dissipation as the 10- Gbit/sec fiber-optic module developments unfold, the total power dissipated by the small packages will remain high enough that thermal design will play a major role.
Another generation of test equipment accompanies another speed generation of fiber-optic modules. Currently, a lab can be stocked with 12-Gbit/sec bit-error-rate testers (BERTs), parallel BERTs, jitter analyzers, optical-to-electrical (OE) converters, and fancy digitizing oscilloscopes that sample fast enough to produce eye-diagrams with 10 psec per division granularity. As in every generation, there will be questions on how to measure certain parameters such as jitter. Measurement of jitter historically has been a topic of much discussion, most notably by Fibre Channel study groups.
With different data rates per optical channel, different data rates per parallel module electrical interface channel, different clocking schemes of data in and out on a module, and many types of integrated functionality, there may be different techniques for measuring jitter, each most appropriate for one type of module. Compare, for example, a serial duplex 10-Gbit/sec module with a parallel 4x2.5-Gbit/sec module. A time-domain jitter measurement done on an oscilloscope capturing an eye-diagram coming from the transmitter of a serial module might be "enough" to characterize the jitter to predict link performance. Looking at jitter on one of four 2.5-Gbit/sec transmitter outputs in a parallel module might not predict link performance, however. The jitter testing of a 4x2.5-Gbit/sec module might require more specialized testing, whereby patterns are sent through the four channels to test for crosstalk effects on jitter.
The economic slowdown does not decrease traffic on the Web. The need for a high-bandwidth communications network still exists. Previous-generation 10-Gbit/sec modules do not fit the economic mold of widespread proliferation into LANs, MANs, and SANs. Next-generation 10-Gbit/sec fiber-optic modules are already appearing on the market for applications in different areas. These modules incorporate new devices and new ICs with high levels of integration.
Several marketing forces are helping to define technically and economically feasible interfaces to these modules. Undoubtedly, several types of modules (e.g., serial, parallel, WDM) will provide value to different segments of the market at different stages in the generation of modules. Specification and design of modules will depend on electrical, technological, thermal, and metric challenges at 10 Gbits/sec.
With the number of companies and individual contributors involved, and with the large number of dollars of investment money spread among these parties, the excitement of the 10-Gbit/sec module generation can be expected to be fast and furious. Many types of modules are needed and will have their place. One can expect to see each type make its mark within 12 to 18 months, with each type vying for as large a market as possible. In the end there will not be any losers, and consumers and users will all benefit.
Aaron Schultz is a principal applications engineer at Quake Technologies Inc. (Ottawa, Ontario). He can be reached via the company's Website, www.quaketech.com.
- "A Comprehensive WAN, LAN, and Very Short Reach (VSR) PMD Solution," by Pat Gilliland, Dipak Patel, Luis Torres, IEEE 802.3ae, July 7, 2000.
- Story of a Faculty Entrepreneur," Engineering Today, Volume XXVI, pp. 15-19, University of California at Santa Barbara, Winter 2001).