Comb lasers target connectivity
Today s computers are rarely stand-alone machines. Instead, they are hooked up to the information super-highway where they exchange data in the form of e-mails, webpages, pictures, music and video.
The transfer of information involves optical fiber networks, which provide medium- to long-reach data communications via wavelength-division multiplexing (WDM) transmission. However, as the reach gets shorter, there is a switch from optical to electrical links, such as interconnects used in mainstream computing applications. Here the electrical link infrastructure is being pushed to its limit, thanks to the proliferation of server farms, high-performance computing and chip-level multiprocessors.
To address the impending data-rate crunch over these short distances, researchers are developing techniques to unite silicon photonics and fiber optics. These technologies should equip silicon with tools to couple, route, (de)multiplex, detect and modulate light for short-reach needs. However, silicon cannot provide the laser. This means that there is a massive potential market for III-V lasers at wavelengths of at least 1.2 µm, which can provide the optical sources for short-reach communication.
Unfortunately it is difficult to make optical systems that can replicate the levels of integration found in silicon ICs. Arrays of VCSELs and, to a lesser extent, edge-emitter lasers are two of today s best options, but their excessive cost and complexity make them impractical for very-high-volume computer applications. So alternatives are needed, such as single Fabry-Pérot quantum-dot broadband comb lasers that produce a collection of equally spaced emission frequencies. We are developing such a device at Innolume, which is specifically designed for short-reach WDM.
The WDM approach that we are pursuing isn t the most popular method for transferring information over short-reach interconnects. Parallel optics approaches are preferred, which are based on directly modulated VCSELs and multiple fibers, waveguides or free-space optics. However, WDM, which involves the coupling, modulating and multiplexing of light from multiple continuous wave lasers into a single fiber or waveguide, is equally capable of satisfying the need for multiple light source integration on a single chip.
WDM technology also has an advantage over VCSEL arrays: it doesn t require coupling to multiple waveguides or fibers. The VCSEL-based approach fails to offer a clear path for the migration of the optics to the IC because the device has to be placed next to the silicon chip and connected electrically. To eliminate inefficient electrical links, the optics must be brought close to, and eventually onto, the processor or memory, which could lead to direct communication between cores in multicore chips.
Our lasers will target deployment in systems delivering data rates of at least 1 Tbit/s, because this will be the likely tipping point for transition from electrical to optical interconnects. Obviously the crossover point is a moving target because electrical engineers will continue to extend this technology s capabilities. There will come a time when the engineers are defeated by power issues at ever higher speeds, such as greater distortion of electrical signals (skin effect), dielectric losses and reflections caused by impedance mismatch. However, higher speeds alone can t guarantee success because compelling cost advantages are also needed to drive the switch to optical signal distribution. This change-over will not be trivial, but a disruptive and expensive transition for the IC, computer, PCB and packaging industries.
We believe that our WDM links should employ signal modulation speeds of 40 Gbit/s or less. Using faster external drivers for these lasers is prohibitively expensive, and employing lower speeds and more channels offers better value for money.
Laser connections
This approach requires optical coupling of the laser to the silicon chip, or the photonic integrated circuit chip in the case of III-V substrates. In our opinion, the best way to do this is to locate the laser assembly off-chip and connect this with an optical fiber/ribbon or integrated waveguides. This is not a cheap solution, but it does deliver manufacturing advantages, as the laser and silicon ICs can be brought together during motherboard or package assembly.
The alternative to this approach is based on the bonding of laser array chips to silicon or silicon photonic ICs. However, IC manufacturers may not favor this cheaper method because they will be reluctant to relinquish chip real estate and will want to avoid making disruptive architectural changes to their highly evolved and very valuable jewel – the CPU. Bonding approaches are restrictive because the single-frequency emission needed can only be produced by distributed feedback and microring laser designs. However, Intel and UCSB have shown that gratings in the silicon waveguides can provide feedback.
Bonding technologies also hamper the laser s operating environment. High-performance chips, such as dual-core CPUs, idle at 30–40 °C but work at 70, 80 or even 100 °C under heavy loads. High temperatures and large temperature swings degrade the device and drive variations in quantum-well laser emission wavelengths. Quantum-dot lasers are immune to this – the Fujitsu spin-out QD Laser Inc recently announced a 1310 nm emitter for passive optical network uplinks that is insensitive to temperatures between –40 and 100 °C.
Regardless of whether the laser chip is mounted externally or onto either the silicon photonics chip or the photonics integrated circuit, it must mimic the performance of an array of distributed feedback lasers producing up to hundreds of channels. This means that each channel should deliver at least 1 mW with very low relative intensity noise to ensure low bit error rates at high modulation speeds. On top of all of this, the laser chip must command a price tag of about a dollar for high-volume IC applications.
The quantum-dot lasers promise to fulfill all of these requirements. They take advantage of the intrinsic variations in quantum-dot sizes, which lead to a broad lasing envelope. By intentionally producing structures of up to 10 layers of quantum dots with different dot sizes in each layer (figure 1), we can make a device with a very wide optical gain that delivers broadband lasing at high power (figure 2). The quantum-dot lasers that we have fabricated include a device with a 75 nm wide emission spectrum and only 3 dB intensity variation.
Our broadband laser does not produce continuous emission over a wide spectrum – such a device would be too noisy for high-speed external modulation. Instead we have built a broadband comb laser, which emits a collection of single Fabry-Pérot lines. The relative intensity noise of each line (longitudinal mode) is the critical parameter, and calculations show that this must be less than 0.3% for 10 Gbit/s transmission with a bit error rate of less than 10–12.
Guaranteeing modal intensity stability in any conventional Fabry-Pérot laser is not just a matter of maintaining overall power stability, as different longitudinal modes compete for optical gain. Noise associated with this, known as mode partitioning, prevented engineers from using quantum-well lasers as a simple comb spectrum source several years ago.
However, we have solved this problem with proprietary quantum-dot engineering. Also, experiments conducted by the Heinrich-Hertz-Institut in Berlin have revealed that individual lines of a comb laser support error-free 10 Gbit/s transmission using external modulation (figure 3).
Our best results to date include 10 mW per channel over 16 channels and more than 1 mW per channel over 100 channels, for channel spacing of less than 50–140 GHz (
