Faster, more frugal, greener VCSELs

BY DIETER BIMBERG FROM THE TECHNICAL UNIVERSITY OF BERLIN, ANDERS LARSSON FROM CHALMERS UNIVERSITY AND ANDREW JOEL FROM IQE

When you hear the phrase "˜heavy industry' your mind might conjure up images of colossal steel-making plants, vast, busy dockyards, or endless rows of gritty smokestacks belching unpleasant and toxic gases aloft. But in the twenty-first century, you might equally well think of massive environmentally controlled data centres lying just a stone's throw from cold-water lakes or other natural chilled water sources, and surrounded by high-latitude evergreen forests. Due to the tremendous amount of energy these data centres now consume, such places of extreme data storage, transfer, and manipulation are our modern-day factories supporting the world's infrastructure, daily operations, and thirst for information and instant communications.

To aid their running, datacentres are being located alongside power stations. And to trim the electricity bills, more of them are cropping up in regions with access to low-cost natural and renewable power, such as Iceland, a country blessed with a rich source of geothermal energy, and Sweden, a country renowned for both natural beauty and cold water Baltic bays ideal for datacentre cooling and low-cost power sourcing.

Throughout this decade and into the next, datacentres will increase in number to try and satisfy the seemingly insatiable demand for data everywhere, ubiquitous connectivity, and unlimited and seamless human and machine interfaces, all believed at near zero cost. This vision is driven by trends towards cloud computing, big data, and the Internet-of-Things.

To try and prevent the energy that datacentres consume from getting out of hand, there is a need to develop more reliable, higher-speed data-transmission links that consume less energy. The traditional medium for supporting data transfer is a length of copper wire in the form of a "˜twisted pair' connector, but this has already been replaced in many datacentres with short-reach, laser-diode-based optical links and interconnects, which combine a higher data capacity with greater bandwidth density, higher transmission efficiency, a more robust operating temperature range, longer reach, and lower costs.

These optical links are being continuously refined, leading to faster and faster aggregated transmission speeds, i.e. increased bandwidths (see Figure 1).

Figure 1: Lane rates are increasing to help to try and meet the insatiable demand for more data, and they could eclipse to 100 Gbit/s by the end of the decade.

A key ingredient in these optical links is the light source. Previously the most common form of this device for longer distances was an edge-emitting laser diode, which employs cleaved crystal facets as reflectors to form the laser cavity, forcing coherent light to be emitted from the edge of the device.

One downside of this architecture is that it must involve the fabrication of discrete devices, with manufacturing and testing only performed on fully processed edge-emitting devices. An additional weakness is that the output from these devices shows an elliptical beam profile, which is tricky and expensive to collimate and to focus into the end of an optical fibre.

Due to these limitations, the vertical-cavity-surface-emitting laser (VCSEL) is a far more attractive class of laser diode for deployment in short-reach optical interconnects for datacentres, server farms, and supercomputers. This class of laser contains an active layer confined by two distributed Bragg reflectors (DBRs), which are formed via the growth of multiple thin epitaxial layers on a semiconductor substrate. With this device geometry, light exits the chip through either the upper or lower DBR reflector, thus in a direction that is vertical (perpendicular) to the surface of the VCSEL (see Figure 2 for top-emitting VCSEL). Thanks to surface emission, the VCSEL can be tested on wafer, which delivers an enormous cut in manufacturing cost and lends itself to the fabrication of laser arrays comprising multiple emitters on a single chip. By using direct butt-coupling, adding a lens to the emitting surface of the VCSEL, or by using a fibre with a built-in lens tip, it is possible to couple the emitted light into a standard multiple-mode optical fibre. In high optical output power or illumination applications, arrays of VCSELs may be used in concert with a microlens array to collimate the emitted light.

Figure 2: A typical GaAs-based VCSEL structure features an active region sandwiched between two multilayer mirrors (DBRs). Carriers are injected through the contacts on the top and bottom of the structure, with light emitting through the top. The beam profile is governed by the width of the oxide aperture.

The speeds (data bit rates) of commercially available optical fibre links are held back by the light source, and presently show maximum bit rates of between 10 Gbit/s and 28 Gbit/s. The maximum operating bit rate of a VCSEL depends on several intrinsic device parameters including the damping of the modulation response, thermal effects, and aspects of device design that determine parasitic resistance and capacitance.

Turbo-charging the VCSEL

To spur the VCSEL to higher speeds, our team of researchers at the Technical University of Berlin (TU Berlin), Germany, at Chalmers University of Technology, Sweden, and at the UK-based epitaxial-wafer manufacturer IQE plc, has developed new device architectures that can increase data transmission speeds. This work first evolved as part of a European Commission-funded programme called VISIT "“ Vertically Integrated Systems for Information Transfer, during the years 2008-2011 .

We have made essential refinements to the VCSEL design (Figure 3), including replacement of the GaAs-based active layer with a strained InGaAs/AlGaAs quantum well structure to provide higher differential gain, and the introduction of separate confinement heterostructures to speed up the transport of carriers and ensure low gain compression. In addition, we trimmed the cavity length to boost optical confinement; cut the number of mirror pairs in the top distributed Bragg reflector, in order to cut reflectivity and ultimately shorten the photon lifetime; and introduced more advanced interface grading and modulation doping schemes in the DBR mirrors, to lower resistance and optical losses. Completing this list of enhancements was the inclusion of multiple oxide layers, the introduction of an undoped substrate to lower capacitance, and the use of an AlAs binary compound in the bottom mirror to reduce the thermal impedance.

Figure 3: Refinements of the VCSEL design have propelled speeds to 50 Gbit/s and beyond.

These changes to the VCSEL structures have impacted the epitaxial deposition process. Growth of the new device structures demands exceptionally tight control over compositional uniformity, maintenance of strain, and highly accurate placement of layers for forming an oxide aperture for current and optical confinement.

Changes to the VCSEL designs have wrought improvements in performance. For VCSELs emitting at 850 nm the modifications to the active region and the cavity have delivered a doubling of differential gain, a 30 percent cut in threshold carrier density, and a 20 percent increase in the optical confinement factor. Carrier transport has also improved, while the diffusion capacitance has decreased, as demonstrated by Chalmers.

A better DBR has resulted from the new design. Differential resistance has fallen by 25 percent in the top mirror, while free carrier absorption has showed little, if any, increase. Meanwhile, in the bottom mirror the switch to the pairing of AlAs and Al_0.12Ga_0.88As has delivered a 40 percent hike in thermal conductivity. Parasitic capacitance has decreased between 30 and 40 percent via the addition of extra oxide layers.

Due to all of these changes, VCSEL speeds have been catapulted to beyond 50 Gbit/s (see Figure 4; from Electronics Letters, August 2013) and to such an extent that the optical interconnect speed is largely limited by the speed of the drive electronics and the optical receiver. Recent work at IBM (Yorktown Heights, NY), using Chalmer's 850 nm VCSELs and drive and receiver electronics developed at IBM, has enabled optical interconnects operating at a channel rate of 64 Gbit/s (Optical Fiber Communication Conference, 2014, paper Th3C.2), which is a world record for VCSEL-based interconnects. IBM has also demonstrated transmission at 51.56 Gbit/s (the InfiniBand HDR data rate) with the VCSEL-based transmitter held at a high temperature of 90°C.

 Figure 4: Data transmission over a VCSEL-based optical interconnect at 50, 55 and 57 Gbit/s.

Measurements of bit-error-rate and eye diagrams for 850 nm VCSEL-based optical interconnects at bit rates from 50 Gbit/s to 57 Gbit/s are shown in Fig. 5. The bit-error-ratio is a measure of the probability that a data bit is erroneously detected, with 10^-12 typically considered "˜error-free' for short-reach optical fibre interconnects. An eye diagram is constructed by overlaying sweeps of different segments of the digital data stream, thus creating an "˜eye' displaying the on and off-levels and the transitions between these.

Boosting efficiency

Fast speeds are highly commendable, but they are not the only metric that matters: the energy dissipation per bit is probably still more critical, for it is this that has the larger influence on the power drawn by the datacentres. This is exactly the focus of the work at TU Berlin, wherein the VCSELs have broken every conceivable VCSEL energy efficiency record during the past several years at both 850 nm and at 980 nm (see Figure 5). By making independently similar changes to the epitaxial structure, including those previously listed, and by adding additional current spreading layers and proprietary changes to the quantum wells and distributed Bragg reflector (DBR) mirrors to increase the differential gain and the heat dissipation, the energy dissipation was decisively lowered in 2013 beyond the level requested by the ITRS road map for 2015.

Figure 5: Energy efficiency (a) in units of femtojoules per bit and temperature (b) both versus error-free bit rate of state-of-the-art infrared VCSELs.

We have strived to excel in this area, and earlier this year during the Photonics West 2014 conference (held during 1-6 February 2014, in San Francisco, CA) TU Berlin announced in partnership with IQE plc that error-free operation at speeds of up to 40 Gbit/s had been accomplished, with a record low energy dissipation (below 108 fJ/bit). This dissipated energy per bit at 40 Gbit/s is at least four times less than any other published result for VCSELs. For this work TU Berlin received the SPIE 2014 Green Photonics Award in Communications, after having previously already received this award in 2012.

In more recent work the 980 nm VCSELs that enabled this record also demonstrated extreme temperature stability during high-speed operation at 46 Gbit/s and temperatures of up to 85°C. These superb results were presented by the TU Berlin group at the IEEE ISLC in Palma de Mallorca, Spain during 6-11 September 2014.

The energy dissipation per bit is not fixed for a VCSEL, but depends on the transmission speed, the operating conditions, the intrinsic VCSEL materials and thus on the epitaxial design, and the other components along with the VCSELs that form a complete optical interconnect, such as the VCSEL driver electronics, the photoreceiver, and the modulation scheme. As demonstrated in 2013 when providing error-free operation at 25 Gbit/s, TU Berlin VCSELs have demonstrated that they can deliver a record-low dissipated energy of 56 fJ/bit. This remains to date the lowest reported value of dissipated energy at error-free operation for any semiconductor laser diode at any wavelength or bit rate, and it was achieved at a current density of just ~10 kA cm-². This result demonstrates the suitability of these devices for application in reliable, sustainable commercial "˜green' optical interconnects.

Emerging new markets

Development of epitaxial processes to produce these highly complex structures, alongside refinements to the VCSEL architecture, will not just led to faster and more efficient lasers for datacentres "“ it will also help to drive the deployment of this class of laser in a wide range of new and emerging industrial, commercial, and consumer applications. Penetration into these new markets will be aided by establishing a European production capability that brings VCSEL manufacturing to a level comparable to LED and CMOS manufacturing. Such efforts are underway, with IQE plc taking part in a € 23 million programme announced this May, entitled VCSEL Pilot Line for IR Illumination, Datacom, and Power Applications (VIDaP).

Partners in this European project, which is funded by the European Commission, include Philips, STMicroelectronics, Sick and Sidel. Together, these firms will bring together existing high-volume production facilities at IQE with key end users to create a consortium delivering an end-to-end production supply chain. This should significantly reduce the cost-per-function for the VCSEL, by reducing GaAs processing costs whilst increasing device performance.

One emerging application that could swell VCSEL sales is gesture recognition "“ this could be used for gaming and non-contact navigation. This class of laser could also be used in depth imaging for 3D vision, using "˜time-of-flight' technology. In addition to these consumer applications, VCSEL deployment is set to rise in industry as chip volumes increase and devices offer greater performance at lower cost (see Figure 6). This surface-emitting laser may be used for highly efficient, digital thermal processing. Examples of this are high efficiency production-line heating, paint curing and ink finishing in commercial printing.

Figure 6: According to Philips, wafer costs should fall by a factor of three between 2012 and 2018, while the cost-per-Watt for high power VCSEL arrays should plummet by a factor of six over the same timeframe.

On top of this, it is possible for VCSELs to find deployment in illumination for IR cameras for security, safety and night vision; ultra-high density magnetic storage using heat-assisted magnetic recording, and cosmetics and healthcare devices. There are clearly many opportunities for the VCSEL, and shipments will rise as deployments increase in data centres, and then extend to other applications. Initially, it will be the speed and efficiency of this class of laser that will drive its uptake, but there will then come a time when it is power density at low cost that is responsible for the mass adoption of the VCSEL across multiple end markets.

Laser priniting is a promising application for high-power VCSELs.

Advances in the manufacturing process will have to underpin this growth of VCSEL volumes. The groundwork is already being laid, as production moves to larger wafers, using processes that are similar to those employed in the LED and silicon CMOS industries. It is only a matter of time, therefore, before the VCSEL becomes a mainstream product serving myriad applications.

The authors wish to express their gratitude to James Lott from the Technical University of Berlin for his contribution to the work and this article.