- Silicon Photonics Poised to Invade Local Area Networks
By integrating optical functions into a microprocessor fabrication process, engineers can achieve the performance of optical links within the existing framework of silicon electronics.
Cary Gunn, Luxtera Inc.
The time for the mainstream adoption of optical links in the local area network (LAN) has arrived.
Brave optical engineers have uttered these words each time the need for more bandwidth appeared to challenge the capability of copper links. They almost always have been wrong. Silicon designers, wielding ever more sophisticated design tools and riding the speed improvements of Moore’s law, have always found solutions for the problems facing mainstream copper links.
But although electrons have reigned supreme since the dawn of digital communications, photons have successfully intruded into their domain. When bandwidths increased to 155 Mb/s, photonics gained its first toehold in the small market of long-distance telecommunications. As bandwidth increased to 622 Mb/s and to 2.5 Gb/s, optical links maintained their ground in telecommunications and expanded into markets that involved shorter ranges, down to metro distances, or into those in which customers were willing to pay more for advantages such as lightweight cabling and immunity to crosstalk.
Primarily, these early wins for photonics exploited a fundamental weakness of electrical links: their poor bandwidth-distance product. Each time the required bandwidth of communications links has increased, the longer-distance markets have fallen to photonics (Figure 1).
Figure 1. Photonics technology penetrates the telecommunications sector more deeply every year. As the required bandwidth increases, photonics is finding markets at shorter distances. The dotted lines indicate the interface between electrical and photonics technologies in 1985, 1995 and 2005.
Today, 10-Gb/s links have become prevalent, and they are well-established in the high-end enterprise backbone. Those same silicon designers who conquered 1 Gb/s in the LAN are having a tough time with 10 Gb/s. After years of effort by many of the best designers, constructing chips using the most advanced processes available, and after spending buckets of money from many of the best venture capitalists, there is little to show.
A tough time
So far, no available electrical products appear poised to dominate the 10-Gb/s LAN. Each electronic approach has a problem with at least one key issue, chief among them being reach and power consumption.
It is safe to assume that there will be some success with the 10 Gigabit Ethernet electrical standard called 10GBase-T, but this standard has serious performance concerns that will limit its market penetration. At 10 Gb/s, the bandwidth-distance product becomes a critical weakness.
Although 1 Gb/s could reach 220 m, enough to satisfy most cable installations, the 10-Gb/s community is aiming for 55 m and is hoping for 100 m, distances that are unsuitable for much of the existing infrastructure. This will require the customer to pull (and possibly calibrate) expensive new cabling, and the chip power dissipation will be uncomfortably above the 10-W level.
For the first time, electrical links seem destined to concede a large portion of the LAN market to photonics.
The 10-Gb/s market is otherwise ready for the mainstream. High port-count switches for 10 Gigabit Ethernet and InfiniBand are offered by vendors, and basic infrastructural circuitry such as 10 Gigabit Ethernet media access controllers is widely available. Most major chip interconnect protocols (such as PCI Express, HyperTransport and InfiniBand) can source data at 10 Gb/s and plan to source data at 20 Gb/s in the near future.
Today, when the 10-Gb/s data generated in these systems is transmitted more than ~15 m, it is done using an optical transceiver. The problem preventing the widespread adoption of photonics technology is that old Achilles’ heel: Optical transceivers are expensive.
A solution is to fabricate optical transceiver chips using the same mainstream silicon processes that make microprocessors. It thus becomes possible to offer optical transceivers at the price of electrical ones, while maintaining the high performance standards of an optical link.
Silicon photonics was first proposed by Richard A. Soref of the US Air Force’s Rome Laboratory at Hanscom Air Force Base in Massachusetts. Through a series of papers published in the mid- to late 1980s, he introduced the concept of silicon as an optical material for low-loss waveguides — and of modulation using electrons and holes.1,2 It took some time for these ideas to take hold. Nevertheless, during the heyday of the late 1990s, companies such as Bookham Inc. of San Jose, Calif., and LightCross Inc. of Monterey Park, Calif., popularized the technology, and academic research in the field skyrocketed.
Unfortunately, as the telecom bubble burst, so did these early attempts at silicon photonics. A major problem with these attempts was that each company maintained a boutique silicon fabrication facility that could make its photonic products but little else. The high cost of running and maintaining this equipment was not justified by the markets addressable by the relatively primitive state of the technology. At that point, modulation at speeds of even 1 Gb/s was not possible, and integration with CMOS transistors was out of the question.
Silicon photonics continued to thrive in universities such as California Institute of Technology in Pasadena, MIT in Cambridge, Mass., Cornell University in Ithaca, N.Y., and the University of California, Los Angeles, and in the research labs of large companies such as Intel Corp. of Santa Clara, Calif., STMicroelectronics of Geneva and IBM Corp. of Armonk, N.Y. Occasionally, these institutions announce new components.
Luxtera Inc.’s CMOS photonics die contains 250,000 transistors and more than 50 optical components. It incorporates multiple 10-Gb/s optical transceivers and employs flip-chip lasers as the radiation source.Precision optics should be handled properly in cleaning. Here, the prism is held on the frosted surfaces.
Close examination, however, reveals that these components usually are constructed in a way that is fundamentally incompatible with integration into the narrow process windows of mainstream fine-linewidth technology. This rules out their integration with CMOS transistors and the opportunity for cost-effective mass production.
Luxtera Inc. of Carlsbad, Calif., and Freescale Semiconductor Inc. of Austin, Texas, are developing silicon photonics with true CMOS compatibility. Luxtera has developed CMOS photonics technology capable of integrating complete optical transceivers within a CMOS die (Figure 2). Such a transceiver is a blend of optical components and high-speed circuitry.
Figure 2. The block diagram shows a 10-Gb/s optical transceiver integrated with CMOS circuitry. Optical signals, high-speed electrical signals and slow electrical connections are represented with red, yellow and blue arrows, respectively. TX PHY: Transmitter PHY Circuit. RX PHY: Receiver PHY Circuit. TIA/LA: Transimpedance Amplifier/Limiting Amplifier. DC TIA: DC Transimpedance Amplifier.
The key electrical functional blocks are 10-Gb/s circuitry that communicates with the host system in which the transceiver resides (the PHY circuit) and equally fast circuitry that interfaces directly with the optics (the PMD circuitry). Because these blocks must pass high-speed data, they are challenging to design and manufacture. Additionally, there is a substantial amount of bias and control circuitry, which is less critical because it does not need to operate at line speed.
The key photonic building blocks are modulators to imprint data on the output of the laser, optical filters to multiplex/demultiplex optical signals to/from a fiber, a holographic lens to efficiently couple laser radiation between the chip and an optical fiber, waveguides to route the radiation between components on the die, and photodetectors to convert the received optical signal to an electrical one.
The path of the data through such a typical transceiver starts with the electrical data that arrives at the transceiver from the host system. In the transceiver diagram, this point is labeled “10 Gb/s In (Electrical).” By the time a 10-Gb/s signal has been transmitted from the host to the transceiver die, it has picked up jitter and crosstalk from nearby electrical signals, and it has been severely attenuated and distorted. The electronic circuitry recovers the signal and its clock, which is used to retime the data, removing the jitter from the incoming signal.
Then the signal is transferred to the transmitter circuitry, which consists primarily of the modulator driver. This circuit is designed to amplify and impedance-match the electrical signal with the modulator electrodes.
Then laser radiation is introduced into the system. The only component that is not monolithically integrated is the laser. Despite recent interesting scientific research in the area of silicon lasers, there is no technology with the appropriate blend of performance and CMOS compatibility to justify integration. Therefore, Luxtera has opted to flip-chip InP-based lasers on top of the die (Figure 3). The output of the 1545-nm surface-emitting laser is coupled into the silicon waveguides with a holographic lens.
Figure 3. In the CMOS photonic die, the output of the lasers is coupled into the silicon waveguides via holographic lenses.
The holographic lens is a diffractive optics structure that mode-matches the laser radiation with a silicon waveguide by focusing the output of the laser to a tiny point of about 0.1 μm2 (Figure 4). Simultaneously, the radiation is turned 90° so that the normally incident radiation is guided along the plane of the chip surface. This coupling is routinely accomplished with less than 3 dB of efficiency across the C-band (1532 to 1562 nm) using only passive alignment.
Figure 4. A scanning electron microscope image shows the holographic lens that mode-matches the laser radiation with a silicon waveguide. The dashed lines represent the core of a single-mode fiber placed on the line normal to the lens. The waveguides are constructed from silicon and are less than 1 μm in each dimension. The ring resonator between the waveguides has a radius of 2.5 μm.
After being coupled into the waveguide, the laser radiation is routed to a Mach-Zehnder interferometer. This device operates by splitting the radiation into two paths, or “arms,” and recombining it after a few millimeters. The incoming electrical signal controls the phase difference of the radiation in each arm, so that when the radiation recombines, it interferes either constructively to produce a “1” or destructively to produce a “0.”
In a system employing wavelength division multiplexing (WDM), the output from a number of modulators is guided into an optical multiplexer, such as an arrayed waveguide grating (Figure 5). If the wavelength of the lasers is chosen correctly, the grating works as a prism in reverse, combining the individual wavelengths into a single waveguide, which is then routed to a holographic lens for coupling to the outgoing transmission fiber.
Figure 5. Integrated control circuitry is used to tune the arrayed waveguide grating for optimal operation.
Electrodes placed along the grating’s waveguides can generate free carriers to tune the device’s refractive index to compensate for thermal drift. A small amount of laser radiation tapped from the output waveguide provides an error signal to ensure alignment between the gratings and the lasers. This operation consumes very little power on chip and allows the die to operate without a thermoelectric cooler, resulting in cost and power savings compared with traditional WDM systems.
Once in the fiber, the signal transmission is very clean. A piece of single-mode fiber has a bandwidth of about 80 THz. (In comparison, Cat 5e cabling has a bandwidth of only about 350 MHz.) Attenuation in fiber is 0.25 dB/km or less. This means that a 300-m span will lose only 2 percent of the transmitted power — after 2 km, only 10 percent is lost. Also, dispersion is negligible at these distances, so there is no need for dispersion-compensation techniques. Furthermore, optical crosstalk between adjacent cabling is nonexistent. Electrical cables, on the other hand, seriously attenuate and disperse the signal, while at the same time picking up crosstalk from adjacent cables.
At the far end of the fiber, another holographic lens couples the laser radiation into the receiver section of the silicon die. The radiation is sent through another arrayed waveguide grating for demultiplexing into individual outputs. Again, this grating is tuned to track the incoming wavelengths. Each output wavelength is directed to a photodetector, where it converts back into a weak electrical signal.
The photodetectors are constructed from single-crystal germanium, which is grown directly on top of the silicon waveguides. These detectors have very low capacitance, which allows the receiver to be designed for enhanced sensitivity. The detector output signal is sent to the receiver circuitry, where it is amplified with a transimpedance amplifier designed for the integrated detectors, and subsequently is sent to a limiting amplifier and a clock recovery circuit.
This fully recovered electrical signal is reclocked, amplified and passed on to the printed circuit board. At this point, the CMOS photonic transceiver has transmitted data over distances sufficient for operation of the LAN, and it has done so using substantially less power than competitive solutions, all from a single silicon chip.
Although CMOS photonics shines for single-channel devices, its true power lies in its ability to scale. By adding wavelengths, the transmission bandwidth down the same optical cable grows by leaps and bounds — and this does not require a process change.
Figure 6. The CMOS photonic die, approximately 8 × 5 mm in size, contains the functionality of two optical transceivers in a fraction of the size. TX PHY: Transmitter PHY Circuit. RX PHY: Receiver PHY Circuit. TIA: Transimpedance Amplifier.
At last year’s Supercomputing Conference in Seattle, Luxtera demonstrated a 40-Gb/s transmitter using four wavelengths, each running at 10 Gb/s (Figure 6). It is working on 100 Gb/s and has engaged with key customers interested in transceivers as fast at 1 Tb/s.
The future for cabled transmission using silicon electrical signals is looking bleak. The era of silicon photonics is just beginning.
Meet the author
Cary Gunn is co-founder and vice president of technology at Luxtera Inc. in Carlsbad, Calif.; e-mail: email@example.com.
1. R.A. Soref and J.P. Lorenzo (June 1986). All-silicon active and passive guided-wave components for λ = 1.3 and 1.6 μm. IEEE J QUANTUMELECT, pp. 873-879.
2. R.A. Soref and B.R. Bennett (January 1987). Electrooptical effects in silicon. IEEE J QUANTUMELECT, pp. 123-129.
MORE FROM PHOTONICS MEDIA