Low-cost interconnects for 10-Gbit/sec data transmission
Higher bandwidth is exposing the weak links between equipment in central and enterprise offices. This 85O-nm VCSEL-based approach, still in the development stages, holds promise for this future market.
Robert A. Marsland
New Focus Inc.
The surge in Internet and other data services is creating a cascade of higher-capacity requirements from the desktop to the backbone and beyond. This wave of increasing demand is exposing numerous bottlenecks in the network. One problem area is the interconnects between network equipment at central or enterprise offices, as routers, switches, and dense wavelength-division multiplexing (DWDM) terminals handle more and more data. The current solution to this problem is to use expensive Synchronous Optical Network (SONET) transmitters and receivers for these intra-office links. As the number of these short-reach interconnects expands, the cost of such solutions is becoming prohibitive.
At the same time, data networks are evolving to 10-Gbit/sec rates. Standards bodies are already evaluating solutions for 10-Gbit/sec Ethernet and Fibre Channel. Expected to emerge within the next two years, these standards will help drive the market for low-cost, 10-Gbit/sec short-reach interconnects. Analysts project billions of dollars in revenue from sales of 10-Gbit/sec data links within a few years. Developers are responding to this emerging market opportunity with a number of solutions, many tailored for a particular segment. The choices start with copper or fiber; with fiber, the options include singlemode or multimode, single-fiber versus multifibers, and single-wavelength or multiwavelengths. As demand in creases, and networks migrate to faster transmission rates, the need for a low-cost, 10-Gbit/sec intra-office interconnect is apparent.
In any discussion of high-performance interconnects, it is important to remember that the beginning and end of the journey are electrical. The typical reason for resorting to a fiber solution is that copper cannot reliably transmit data over the required distance and meet electromagnetic compatibility standards. At shorter distances, however, the advantage switches back to copper. For example, an integrated 10-Gbit/sec solution has reached 21 ft over a single 0.19-inch coaxial cable.1 Multichannel transmission or multilevel signaling at reduced baud rates can potentially extend this distance.
The difficulties associated with high-speed, multichannel, electronic data transmission that fiber-optic solutions may ameliorate include crosstalk, echo, and channel-bandwidth limitations. With copper, these issues are resolved using digital signal processing (DSP) at either end of the link. The speed of the available DSP technology, however, limits these solutions. Even with DSP, there is a practical limitation to the distance data can be transported economically over copper wires. Based on presentations made to the Institute of Electrical and Electronics Engineers' 10-Gbit/sec Ethernet task force (IEEE-802.3ae), this limit is probably <100 m, although electronic designers are constantly challenging this limit.
Once the decision is made to use fiber-optic cable instead of copper, the next choice is which fiber type. Singlemode fiber has been used almost exclusively for long-distance communication (>2 km), while local area networks (LANs), including campus networks, were built with multimode fiber. This approach is due primarily to the expense of singlemode transceivers, but the high prices and difficulties associated with singlemode interconnections also play a role in decisions to use multimode fiber.
The larger-core 62/125-micron fiber, which has greater light-collecting ability but the unfortunate property of less than 200-MHz-km modal bandwidth at 850 nm, is the most popular multimode choice. Modal bandwidth measures the velocity dispersion of individual transverse modes propagating down the fiber core. Longer propagation distances allow the different modes to separate in time. This separation can be visualized as a narrow pulse, composed of many transverse modes, introduced at the input of the fiber, then spreading out into a broad pulse packet as the different modes travel at different velocities. A consequence of limited modal bandwidth, along with all other bandwidth limitations, is intersymbol interference (ISI).
Given these multimode limitations, the inclination is to use singlemode fiber. However, the only low-loss, low-cost singlemode fiber available requires the use of laser wavelengths >1,100 nm for robust transmission. Multigigabit transmission at 850 nm over singlemode fiber at multikilometer distances has been demonstrated in a lab.2 Singlemode transceivers continue to cost more because of the fiber's requirement for long-wavelength lasers and <1-micron alignment.
At least two fiber manufacturers, Corning and Lucent Technologies, have responded to this situation with improved high-modal-bandwidth fiber. The Lucent fiber, called LazrSPEED, can provide an order of magnitude improvement to over 2,000 MHz-km. Corning has announced similar results with its Infinicore fiber. When combined with a narrow linewidth laser, the LazrSPEED 50/125-micron fiber has been demonstrated to carry 10-Gbit/sec data over distances >1 km.3 These results show that 850-nm lasers combined with this new multimode fiber will continue to provide a natural migration path for data links within the enterprise.
Another approach to creating more interconnect bandwidth is to increase the number of parallel fiber channels. If some economies can be achieved with mass termination of the optics and parallel electronics, then it is reasonable to expect that a parallel optic transceiver could be less expensive than separately purchased lower-speed links. As others have pointed out, it is clearly the case that economies can be found in the mass-termination of optical transceivers. However, parallel fiber and connectors have not achieved prices below the separately purchased parts. This pricing is due to a lower volume application of parallel fiber and the fact that there is no economy in manufacturing a parallel ribbon of 12 fibers versus 12 individual fibers except in the protective sheathing.
The parallel fiber connector is particularly difficult since it is essentially an over-constrained mechanical problem: 12 fibers must simultaneously make face contact with 12 other fibers. Contrast this with the single fiber connector that can hardly fail to make contact with another fiber inserted in the same sleeve under spring force. For now, the expense of the connector and cable completely override any advantages gained through parallel termination at the transceiver.
Cost considerations aside, multifiber and multiwavelength approaches both create the following difficulty: The serial-data stream must be split into a number of independently clocked serial-data streams. In the packetized data environment, the result is proportionally relaxed timing requirements on the individual data streams. The data streams can be realigned at the receiving end of the link using one of several schemes now under study within IEEE-802.3ae. In SONET, however, it is usually necessary to transmit clock information on the data stream. In this situation, the timing requirements are not relaxed regardless of the data rate on the individual channels, which eliminates much of the desired benefit in running at reduced rates.
Finally, it is possible to increase fiber capacity in the data network by increasing the number of wavelength channels transmitted over the fiber, technology that is now widely used across the nation's telecommunications infrastructure. The wavelength channels are combined and separated using optical multiplexers and demultiplexers.
Several companies are pursuing an approach where the astronomical cost of wavelength-division multiplexing (WDM) is mitigated by using a small number of widely separated wavelengths. One such company, Agilent Technologies, refers to its technology, which uses four wavelength channels near 1,300 nm, as wide wavelength-division multiplexing (WWDM). This powerful technique is potentially compatible with the installed base of multimode and singlemode fiber over link distances that will satisfy most data-networking applications. The downside of this approach is similar to that of multiple fibers; the data stream is broken up and transmitted over separate independently clocked channels. The success of WWDM will probably depend on the extent to which the complex optical assembly can be produced at low cost.
The most obvious approach to low-cost 10-Gbit/sec data communication is vertical-cavity surface-emitting laser (VCSEL)-based serial transmission over multimode fiber. This approach is inherently low cost, but it also introduces three potentially significant issues-the package, the VCSEL, and the electronics. These risks, however, seem manageable compared to the drawbacks of the competing approaches.
For 10-Gbit/sec serial transmission, it was apparent that there was a need for a low-cost, optoelectronic package. The package design had to satisfy the following requirements:
- Accommodate both optics and electronics (like a "butterfly" package without the complexity).
- Compatible with complete burn-in and test prior to fiber attachment (like a transistor outline style package).
- Hermetic, at least as an option.
- Electrical return loss >15 dB up to 15 GHz.
- Acceptable electromagnetic interference.
It became clear that what was needed was a standard hermetic package from the radio-frequency (RF) electronics industry. These packages are available at moderately low cost; the only disadvantage is that there is no provision for optical fiber. In addition, a surface-emitting laser would have to emit through the lid requiring the package to be tipped on its edge. Fortunately, 3-D electromagnetic simulation indicated that a differential interface could provide the required RF performance in this configuration.
The design of the standard microwave package was modified to move all of the leads to one side (see Figure 1), requiring modifications to the metal package body, the ceramic substrate, and the lead frame. But even with these changes, the package is no more difficult to manufacture than the standard microwave version. The advantage of the differential interface is that ground-plane discontinuities pose less of a problem.
The backdrop of this effort has always been the VCSEL performance, which, for oxide-confined VCSELs, has been adequate. Unlike edge-emitting lasers, VCSELs do not exhibit drastic ringing at the laser's relaxation oscillation frequency. That means the signaling rate can approach the VCSEL relaxation oscillation frequency in the 10- to 15-GHz regime. The real question on the VCSEL side has been the manufacturability and reliability of the oxide-confined VCSEL, which provides this performance at the desired 10-mA drive levels.
Honeywell has published reliability data on both its proton-implant and oxide-confined VCSEL lasers.4,5 The oxide-confined VCSELs were found to have larger activation energy and lower reliability at temperatures above 125°C. At temperatures below 125°C, however, these lasers were better than or equal to the proton-implanted devices. Since the drive level used to achieve 10-Gbit/sec performance is within 20% of the drive current used in the Honeywell study, it was concluded that the reliability was not a major risk. Early results from further studies at Honeywell indicate that drive currents as high as 20 mA will pose no reliability problems.
In recent months, several suppliers have introduced oxide VCSELs for 2.5-Gbit/sec data transmission. These devices are nearly identical to those used in this study. With reduced bond-pad capacitance and a relaxed extinction-ratio specification, these devices can perform at 10 Gbits/sec.
One criticism often leveled at high-speed serial implementations is that the electronics are too power-hungry and expensive. Electronics designed for long-haul SONET OC-192 (10-Gbit/sec) applications confirm this claim.
To see if there was a way around this dilemma, a simple feasibility calculation based on the parameters used at a commercial gallium arsenide (GaAs) heterojunction bipolar transistor (HBT) foundry-focused on integrated-circuit development for the wireless market-was used. By using a process aimed at economical wireless devices, the hope was to beat the economics of the high-performance solutions. It was determined that the moderate-performance HBT should be able to achieve -17-dBm sensitivity following the procedure of Smith and Personik.6 This sensitivity is required to operate a Class I eye-safe 850-nm data link that provides a 5-dB window for transmitter-coupled power, 6 dB for cable loss, path penalties, and margin, with roughly 1-dB left over.
Using the GaAs HBT parameters, a receiver circuit was designed that consumes 1 W, while providing two gain-controlled differential 125-mVp-p outputs. Similarly, a transmitter was designed that consumes 0.25 W. Designs based on silicon germanium (SiGe) are expected to consume even less power. The numbers, however, are already quite good compared to existing 1-Gbit/sec designs.
The complete link then consisted of receiver and transmitter chips packaged together with the photodiode and VCSEL elements in the new package. The transmitter module contained the VCSEL driver chip, VCSEL, monitor photodiode, and bypass capacitors for the power supply. The receiver module contained the receiver chip, photodiode, filter capacitors for the DC-restore loop, and bypass capacitors for the power supply. The links were tested by connecting a data generator to the transmitter module, and the error detector to the receiver, with 1-m coaxial cables. The units were fiber-coupled using fine positioning stages rather than laser welding to allow study of various launch conditions. The fiber-optic input to the receiver was connected to the transmitter fiber output via 2, 50, 75, and 125 m of fiber plus a fiber-optic attenuator. The attenuation was adjusted to achieve several bit-error rates in the 10-9 to 10-11 range; then data was extrapolated to find the 10-12 bit-error power. This procedure was repeated for each fiber length.
It was surprising to find little dependence on fiber lengths out to 100 m; the eye-diagram actually improved with length (see Figure 2). The improved eye is due to the unintentional pre-emphasis added by the transmitter because of the slight VCSEL overshoot. As the fiber length increases, the decreasing fiber bandwidth tends to compensate the peaked response of the transmitter producing a cleaner eye, which could also explain the lack of sensitivity to fiber length. It should be noted that the sensitivities are actually closer to what is normally called "stressed" receiver sensitivity because of the coaxial cables, transmitter extinction ratio, and transmitter edge speeds used in evaluating the receiver. As the transmitter is improved and better 850-nm test equipment becomes available, the limits of the receiver can be probed further.
In the coming months, the numerous approaches being considered for low-cost 10-Gbit/sec interconnects will begin to be sorted out. While serial 1,310-nm solutions appear to be the strong candidates -at the moment-for distances of 2 km or more, a serial 850-nm VCSEL-based approach offers a promising low-cost solution for lengths of 300 m or less.7
Robert A. Marsland is vice president of Focused Research Inc. (Madison, WI), a division of New Focus Inc. (Santa Clara, CA). Robert S. Williamson III, Grant R. Emmel, Dominik Hoffmann, and Renée Nesnidal, engineers at Focused Research, Terri Dooley, an administrator at the same division, and Herman Chui, product-line manager in the telecom division also contributed to this article. The author would like to acknowledge Mike Deger strom, Gregg Fokken, Dan Schwab, and Barry Gilbert of the Mayo Foundation Special Purpose Processor Development Group (SPPDG) for their contribution to this work.
- Walker-RC; Kuo-Chiang-Hsieh; Knotts-TA; Chu-Sun-Yen, "A 10 Gb/s Si-bipolar TX/RX chipset for computer data transmission," 1998 IEEE International Solid-State Circuits Conference. Digest of Technical Papers, ISSCC. IEEE, New York, NY; 1998; 504 pp. p.302-3.
- Schnitzer-P; Jager-R; Grabherr-M; Wiedenmann-D; Mederer-F; Ebeling-KJ, "GaAs VCSELs for biased and bias-free multi-Gb/s data transmission over 4.3 km standard 1300 nm single-mode fiber," Conference Proceedings. LEOS '98. 11th Annual Meeting. IEEE, Piscataway, NJ; 1998, vol.1, pp.164-5.
- Giaretta-G, Michalzik-R, Kolesar-P, "New MMF, how far can we go?" presented at the IEEE-802.3 High Speed Study Group, November 1999 Plenary Week meeting, Nov. 9-10, 1999, Kauai, HI http://gro uper.ieee.org/groups/802/3/10 G_study/public/nov99/giaretta_1_1199.pdf.
- Tatum-JA; Guenter-JK; Johnson-RH, "Manufacturability of VCSEL components and VCSEL products," LEOS '98. 11th Annual Meeting. IEEE Lasers and Electro-Optics Society 1998 Annual Meeting. IEEE, Piscataway, NJ; 1998; vol. 2, pp. 409-10.
- Hawthorne, Robert A., Guenter, James K., Granville, David N. "Reliability Study of 850 nm VCSELs for Data Communications," 1996 IEEE International Reliability Physics Proceedings. 34th Annual. IEEE, New York, NY; 1996; pp.203-10.
- R.G. Smith, S.D. Personik, "Receiver design for optical fiber communication systems," Semiconductor devices for optical communication, Series: Topics in Applied Physics, H. Kressel editor, vol. 39, Springer-Verlag, 1979 Berlin, pp. 117-119.
- Informal IEEE-802.3ae PMD voting on preferred solution for 10 G Ethernet.