Jim A. Tatum Honeywell Sensing and Control
Kenneth P. Jackson IBM Microelectronics
The market demand for high-speed fiber-optic interconnects has been fueled by the ever-increasing need for bandwidth in applications such as storage area networks (SANs) and Ethernet server applications. The speed of these systems has been increasing rapidly-they now are capable of operating in excess of 1.25 Gbits/sec. Vertical-cavity surface-emitting lasers (VCSELs) are the optical source of choice for these applications.
VCSELs will meet performance needs as speeds progress to 2.5 gigabaud (GBd) and beyond. For example, serial data links running at 2.5 GBd and parallel optical links operating at aggregate data rates in excess of 10 Gbits/sec are available, and serial links at 10 GBd also have been demonstrated. We'll explore the use of VCSELs in high-speed data interconnects and address some of the issues relating to standardization of these optical data links.
VCSELs are a relatively new class of semiconductor laser that emits light perpendicular to the wafer. VCSELs have been demonstrated in several material systems but primarily have been commercialized in the infrared using aluminum gallium arsenide (AlGaAs), emitting in the 750- to 860-nm range. Like most semiconductor lasers, the active region consists of multiple quantum wells, but unlike edge-emitting lasers, the mirrors are formed during epitaxial growth using distributed Bragg reflectors (DBR). Because of the complicated epitaxial structure and the requirements of the material, commercial VCSELs are generally grown by metal-organic chemical vapor deposition (MOCVD).
Honeywell introduced the VCSEL as a commercially available product in 1996, and it is now the most prevalent laser source for short-wavelength, high-speed data communications. Previous systems relied on the use of 780-nm edge-emitting lasers, found in high-volume applications such as CD players. When compared to today's VCSELs, traditional edge-emitting lasers have several drawbacks for use in high-speed data communications.
In data-communications systems, reliability and integrity of the optical link is of paramount importance. Properly designed VCSELs have demonstrated remarkable reliability statistics, and they continue to evolve. As an example, the evolution of the reliability statistics for Honeywell VCSELs is shown in Figure 1. The first reliability report for VCSELs was published in 1996,1 and an updated version appeared in a subsequent Honeywell application note in 1997.2 A third unpublished reliability study of VCSELs was conducted in 1998 (also shown in Figure 1). The dotted line in Figure 1 represents an ongoing 1999 reliability study and indicates the estimated line for VCSELs on current evaluation. Even at high-stress conditions, it is difficult to generate enough failures to produce meaningful reliability statistics.
The continued improvement in VCSEL reliability arises from increased wafer-level uniformity and control of critical fabrication parameters. Under nominal operating conditions of 40°C ambient temperature and 10-mA average current, the time to failure for 0.1% of the VCSEL population is in excess of 1 million hours. This represents a more than thousand-fold increase compared to edge-emitting technologies. The results of the 1999 reliability study will be described in more detail in a future publication.3
The long-term degradation mechanisms in edge-emitting lasers are typically the formation of dark line defects or catastrophic optical damage to the laser facet. In VCSELs, the long-term degradation mechanism is thought to arise from migration of impurities in the semiconductor material. Dark line defects have only been observed in VCSELs when the current density is extremely high and well outside the operating regime. This difference in failure mechanisms can affect the integrity of the optical link in different ways. It is well known that degradation in edge-emitting lasers can lead to self-pulsation effects, which in turn leads to elevated bit-error rates. Self-pulsation has not been observed in VCSELs subjected to long-term aging at Honeywell.
In addition to static reliability testing, the reliability of VCSELs in data-communication applications has also been studied. Measurements of the optical waveform, the optical eye diagram, beam divergence, deterministic jitter, and the relative intensity noise have been performed before and after accelerated aging. The results from 10 wafers studied are summarized in Table 1. No degradation in the performance characteristics for data communications was observed in this study. The details of this investigation are given in reference .
Suitability for use in higher-speed data-communications systems has also driven the adoption of VCSELs. The relaxation oscillation frequency (ROF) observed in edge-emitting lasers can be as low as 1.5 GHz, which limits the use of these devices in high-speed applications. Because of a reduced photon lifetime, the ROF for VCSELs is typically in excess of several gigahertz and is critically damped, enabling them to be used in higher-speed data communications.
VCSELs can also mitigate the effects of fiber modal bandwidth on the performance of optical communications systems by launching appropriate modal profiles into the fiber. Because of the narrow angular emission and circular symmetry of the optical beam, the appropriate launch condition for the laser light can be generated with simple molded optics. An improper launch condition can decrease the bandwidth of the optical fiber and lead to bit errors caused by excessive intersymbol interference and increased jitter. A more detailed description of the effects of fiber modal bandwidth on gigabit systems can be found in reference . Standardization of the launch condition for laser light into multimode optical fiber is under development within the Telecommunications Industry Association (TIA). The main issue being addressed by this group is the repeatable, accurate measurement of the fiber modal-power distribution; standardization is expected in the next several years.
Like most light-emitting diodes (LEDs), VCSELs emit light perpendicular to the wafer, which allows them to be tested in wafer form. The entire electro-optic characteristics of a VCSEL wafer can be mapped before forming individual devices, which allows automated die sorting and leads to application-specific sorting of VCSELs from the wafer. The vertical emission from the VCSEL also enables the use of traditional LED-style packages for use with VCSELs, such as surface mount and plastic encapsulated leadframes.
As VCSEL technology has matured, other manufacturability aspects have become evident. An example is the evolution of VCSELs in high-volume manufacturing as shown in Figure 2. The lasing wavelength, which is a good measure of material growth uniformity, is mapped for a 3-inch VCSEL wafer. The uniformity of VCSEL growth has increased substantially from 1996 and has led to a more reliable and cost-effective product.
The ability to leverage high-volume LED-style packaging has also helped in the adoption of VCSELs in data communications. Traditional TO46-style packaging is most commonly used, and other packages taking advantage of the vertical emission are also available. The migration to arrays of one- and two-dimensional emitters is readily achievable with VCSELs because of the wafer-level processing. One-dimensional arrays, with ribbon optical fiber, are at trac tive for "outside-the-box" communications such as server networks. Two-dimensional arrays are being considered for free-space optical interconnects, especially in communications "inside-the-box" between microprocessors and memory, where the electrical interconnections are now the bottleneck.
Finally, the adoption of VCSELs for data communications has been driven by the decreasing availability of low-cost edge-emitting lasers operating in the 770- to 860-nm range. The optical data-storage industry has been moving to shorter wavelength lasers to increase the storage density on DVDs and CDs. This, along with the previously mentioned advantages, has made VCSELs the optical source of choice for gigabit-data communications links.
Two paradigms of engineering new products and systems are to reduce size and increase operating speed. The explosive growth of the Internet has fueled the need for optical transceivers operating at 1.063 GBd and 1.25 GBd for Fibre Channel and Gigabit Ethernet, respectively. Transceivers based on industry standard 1x9 electrical pin configurations and hot-pluggable gigabit interface converters (GBICs) have been the most common package configuration, with the duplex SC fiber connector being the most common optical interface. The need to place more optical interconnects on a single board has driven the introduction of new optical transceivers, known as small-form-factor (SFF) transceivers. These transceivers utilize a variety of new fiber optic interfaces, such as the LC, MT-RJ, and VF-45. Manufacturers have announced multisource agreements for the functionality and size of these transceivers, and pluggable versions will be offered in the next few months. The Photo illustrates a variety of optical transceivers using VCSELs that are designed to meet both increased speed and decreased size requirements.
The demand for higher data-rate links, operating from 2.5 to 10 GBd, is increasing as well. New methods in optical-standards specifications are also beginning to emerge that make data-communications systems easier to design and implement. For example, ANSI X3.T11 (Fibre Channel) has adopted approaches based on a minimum swing in optical power between an optical one and optical zero (peak-to-peak power), thus eliminating the optical extinction-ratio minimum. The purpose of the original extinction-ratio specification was twofold. First, to ensure the integrity of optical links that were based on DC coupling, where the extinction ratio strongly influenced the receiver sensitivity. Second, the extinction ratio was assumed when measuring the optical signal, and average power could be used in link diagnostics.
Present-day optical-test equipment and the current design approaches for optical receivers have mitigated the need for high extinction-ratio optical sources. Most photoreceivers and transimpedance amplifiers are AC coupled and are largely insensitive to the average optical power. In addition, when designing an optical transmitter, one of the more difficult problems is balancing the need to maintain a certain minimum optical-extinction ratio along with a minimal amount of deterministic jitter. The elimination of the extinction-ratio requirement will greatly simplify the design and manufacturing constraints for higher-speed optical systems. FC-PI'98 will be the first standard based on the minimum optical-power swing (referred to as the optical modulation amplitude [OMA]), and approval is expected this year. Other standards organizations such as the IEEE 802 and ATM Forum may incorporate this method in the future.
A study group has been formed under the framework of the IEEE 802.3 LAN standards organization to investigate the viability of Ethernet operating at 10 GBd. Several methods have been proposed for the physical layer. These include (a) parallel optical interconnects, (b) serial interconnection, and (c) coarse wavelength-division multiplexing. The High Speed Study Group (HSSG) has met on a regular basis and expects to adopt a methodology this year.
The parallel optical-interconnect approach has been standardized by ANSI for speeds up to 6.4 GBd, and transceiver modules are now available. Among the advantages of the parallel approach is the lowered per-channel data rate and the reduced latency associated with the serializer and deserializer functions. The link distance will be limited by the amount of skew in the fiber ribbon cable and associated drive and receive electronics. The packaging configuration for the VCSEL and detector arrays has evolved to be a lithographically defined array placed on a copper heat sink. Electrical connection is made to the arrays using TAB bonding of a flex circuit, which provides a testable assembly that is placed into a MTP-style fiber connector assembly. The principal disadvantage to the parallel interconnect approach has been the difficulty in meeting the standards for maximum eye-safe power. It is likely that this year the IEC laser safety standard, IEC 825-1, will be amended to increase the maximum power level by 70%, which will greatly relax the power constraints on 850-nm VCSELs.
The serial approach is being driven by the introduction of multimode optical fiber optimized for short-wavelength lasers. The link length of short-wavelength gigabit systems is limited by the bandwidth of the optical fiber, which was previously as low as 160 MHz-km. New multimode fiber with bandwidth in excess of 2 GHz-km at 850 nm recently has been introduced to the market, which should allow link lengths of 300 m or more for serial short-wavelength links running at 10 GBd.
An eye diagram of a short-wavelength VCSEL being modulated at 10 GBd is shown in Figure 3. The optical bandwidth of the laser, the cost of the drive electronics, and the requirement of a new fiber installation are the principal drawbacks to this approach. While the intrinsic speed of the VCSEL can accommodate the higher data rates, the electrical parasitics leading to the VCSEL must be carefully considered. Several approaches have been tested, including placing the VCSEL and laser driver in a chip-on-board fashion and in close proximity, then using ribbon-bond wire to connect them. Another approach has been to directly mount the VCSEL on the laser driver, using either flip chip techniques or bond wires as the interconnect. The laser drivers are being optimized for the variable impedance of the VCSEL to maintain integrity of the electrical signal.
The final method being proposed is use of a coarse wavelength-division multiplexed (CWDM) transceiver. The proposed CWDM systems consist of four channels spaced approximately 10 nm apart, with each channel operating at 2.5 to 3.125 GBd. They will work at short distances with currently installed multimode optical fiber, with extended lengths attainable in the new higher-bandwidth optical fiber. In principle, this approach is suitable for both short- and long-wavelength optical links.
Packaging of the CWDM optical components is still in the developmental stages. The difficult part of the packaging is combining and separating the individual wavelengths at the transmitter and receiver ends, respectively. Definition of the precise wavelengths, and the tolerances on these wavelengths, will need to be carefully considered for interoperability in a standards-based application. Since the precise wavelengths of the system will be predetermined, it is not practical to epitaxially define a single VCSEL array. Therefore, the emitters will have to be individually aligned, which will significantly complicate the construction of optical subassemblies intended for low-cost data communications.
One of the problems facing all the optical and electrical systems proposed for 10-Gigabit Ethernet is encoding the data. Traditional 8B/10B encoding schemes increase the required bandwidth by approximately 25% but result in significantly less complicated receiver designs, which ultimately lead to lower costs and smaller sizes through component integration. Finally, as the lines between data communications and telecommunications become less defined, more universal protocols will need to be developed. The data-encoding me thod is still under debate.
High-speed fiber-optic data-communication systems will continue to grow in support of the demand for information technology. Higher speeds, smaller sizes and increased functionality will continue to be the direction of growth for fiber-optic data-communication links, and VCSELs will continue to be the optical source of choice.
Another advantage to using VCSEL sources in data communication is the potential to address many different signal rates and protocols with a single source. For example, auto-negotiation of a link operating from 10 MBd to 2.5 GBd is possible with a single VCSEL source operating a diverse range of data protocols. In addition, VCSELs emitting in the 1300-nm range are in the development stage at several companies, with production expected in 2001. VCSELs at 1300 nm should provide many of the performance advantages of VCSELs found in the 850-nm regime, and will eventually replace the last remaining edge-emitting lasers in data communication links.
- J. K. Guenter, R. A. Hawthorne, D. N. Granville, M. K. Hibbs-Brenner and R.A. Morgan, "Reliability of Proton-Implanted VCSELs for Data Communications," Proc. SPIE vol. 2683, 1996.
- "Reliability of Honeywell 850nm VCSEL", Honeywell application note, 1997.
- J. A. Tatum, A. Clark, J. K. Guenter, R. A. Hawthorne and R. H. Johnson "Commercialization of Honeywell's VCSEL Technology," to be published in SPIE proceedings 3946, 2000.
- P. Pepeljugoski, J. Abbott and J. A. Tatum, "Effect of Launch Conditions on Power Penalties in Gigabit Links," NIST Symposium on Fiber Optic Measurements, Boulder, CO, 1998.
Jim A. Tatum is manager of VCSEL engineering for Honeywell Sensing and Control (Richardson, TX). He can be reached at (972) 470-4572 or firstname.lastname@example.org eywell.com. Kenneth P. Jackson is a senior engineer for the IBM Optical Interconnect Technology Group (Rochester, MN). He can be reached at (507) 253-6949 or email@example.com.