Abolishing Copper Interconnects
Rocketing demands for data transfer are signaling a switch from copper interconnects to those based on optical fiber. But these new links will only receive widespread adoption when they are paired with ultra-high-speed sources, such as our VCSELs that combine record-breaking modulation speeds with high temperature operation, say Werner Hofmann and Dieter Bimberg from TU-Berlin.
Today’s interconnects, like those of yesteryear, are predominantly made from copper. This material is readily available and easy to process, but it is far from ideal for making data transmission links. Speeds are limited due to high levels of damping that suppress signal-to-noise ratios, and crosstalk plagues interconnects at higher frequencies.
One can get around these weaknesses by moving to higher signal levels and multiple parallel links. This has happened with CPUs, which incorporate more and more pins and use bigger and bigger heat-sinks for cooling. But this approach has its limits, and the days of copperbased technology are vanishing fast for the supercomputer.
In this arena, every new generation of machine must deliver a tremendous hike in computational speed, while costing little more than its predecessor and consuming hardly any more energy. To meet these requirements, the PetaFlop Supercomputer that IBM unveiled in 2008 incorporated 48,000 optical links. And according to this industry’s roadmap, the ExaFlop mainframes of the next decade will feature a staggering 320 million optical interconnects, each performing at far greater levels than those in the 2008 machine: Data transmission through each link will have to be five times higher, energy consumption 50 times less, and the price lower by a factor of 400. Improvements in energy consumption must be this great, because the vast majority of energy consumed by these computers takes place in interconnects.
The delay in the uptake optical interconnects is partly due to strong competition from copper links, which are cheap to produce. Cash has been tight in the optical interconnect industry, hampering investment in new technologies, and big players in the optics industry have focused on markets offering a better short-term return on investment. But that tide is now turning. There is growing realization that optical interconnect technologies are needed, and companies are injecting large sums of cash into their R&D departments to make this happen.
Although funding is flowing, success is never instantaneous. This explains the dilemma that is resulting from the postponed transition from copper to optics. Concerns are so severe that they have led to the publication of articles with headlines such as The End of Supercomputing. Such stories have appeared because it is possible to show that all of the world’s electrical energy generation could go into just bits and bytes, by extrapolating current growth in bandwidth demand, using today´s values for energy consumption per transmitted bit for supercomputers and data-centers.
The ill informed may try to brush aside this looming catastrophe, arguing that humanity doesn’t need supercomputers. What they don’t know is that the Internet is run by large data-centers – when you press a single button for an Internet search, 300 computers scurry away to deliver rapid suggestions. How can these computers communicate, before transmitting the information to the end user? That’s right! By interconnects. Thousands of them.
All theses computers are built around microelectronics, which yield circuits that are far more compact than those built with integrated optics, due to the far shorter wavelength of electrons than photons. Billions and billions of dollars have been poured into the development of silicon microelectronics, but all that funding cannot overcome a fundamental limitation: Optics is much more energy-efficient for interconnects once the bandwidth-length product exceeds a certain value, which is now considered to be relatively low. This limitation of microelectronics is propelling the world towards optical interconnects while one key issue is still to be resolved: How to realize the electrical to optical transition? The existing technologies employed in long haul links are too expensive and energy hungry to be suitable for use in short and ultra-short optical links.
Turning to VCSELs
The most suitable, simplest high-speed optical source is the directly modulated laser. If this is to combine excellent beam quality with low cost and frugal use of energy, the VCSEL has to be the technology of choice. Used in short optical interconnects, VCSELs must excel on three fronts: They must deliver high serial bandwidths, allow dense packaging, and operate without cooling.
System design rules, such as Amdahl’s law, can reveal the required serial bandwidth. Applying this law shows that it is possible to avoid bottlenecks in a supercomputer by matching computational power to interconnect bandwidth and memory capacity. However, simply adding more links to increase bandwidth can lead to penalties, in terms of complexity and cost. What’s more, with one particular technology, there is a limit to the number of links that can be connected. In 2011, these considerations led system designers from Google to state that 40 Gbit/s will be the desired bandwidth for their next-generation data-centers.
To enable system scalability, the optical chips employed in these links must be housed in very dense packaging. One of the strengths of the VCSEL is its small footprint – it is an order of magnitude smaller than it edge-emitting cousins. It is also compatible with a compact hybrid package design, such as that adopted in the IBM TERABUS project. Here, a tiny package resulted from placing bottom-emitting devices with electrical fanout on one side, and using the other side for optics. Such a design requires a transparent substrate, thereby preventing the use of 850 nm VCSELs, the standard source for multi-mode fiber links providing data transmission over hundreds of meters.
Keys to successful VCSEL research include a high yield full-wafer fabrication process, systematic design variation and automated evaluation
980 nm VCSELs are suitable, due to transparency of the GaAs substrate at this wavelength, and they have one big benefit over their shorter wavelength siblings: They enable the construction of distributed Bragg mirrors with binary layers, rather than more complex ternary variants that degrade the thermal conductivity of the chip. Thanks to this advantage, devices built with binary mirrors can operate at much higher ambient temperatures, such as those found when lasers are placed in uncooled dense arrays or integrated on top of a high-performance silicon CPU or memory.
Record-performing VCSEL devices
Taking these considerations into account, out team of researchers at the Center of Nanophotonics at the Technical University of Berlin has focused on 980-nm VCSEL devices of ultimate bandwidth under direct modulation. The VCSEL is a well-established device manufactured in very high-volumes, so if our efforts are to be beneficial outside the lab, they must employ processes used in industry. To that end, we fabricate full three-inch VCSEL wafers with a very high production yield. Our lines of enquiry for improving VCSEL performance are transferred into the mask layout, which features systematically varying device design parameters. After processing in our own class-10 cleanroom dedicated to the development of III-Vs, a robot automatically measures the performance characteristics of thousands of devices.
We use home-built software to handle the vast amount of data generated by these measurements. This enables us to unveil the best chip design from statistical evaluation of measured data. The winning formula is verified with data-transmission and system experiments.
A great strength of this approach is that it enables rapid progress, by exploiting advantages of techniques used in industry and academia. We are able to carry out every step in-house, from simulation to design, epitaxial growth, device processing and a wide range of characterization techniques, including high-speed data transmission experiments. This is an enviable suite of facilities that many groups do not have.
Efforts in our lab have recently led to tremendous improvements in the performance of 980 nm VCSELs over a wide temperature range. Devices fabricated from a single VCSEL wafer have broken the record for the highest speed, enabled 40 Gbit/s operation at temperatures up to 40 °C, and set a new benchmark for energy efficiency at bit rates beyond 30 Gbit/s (see Figures 1 and 2).
Fig 1. Summary of the results achieved on high-speed temperatures table VCSELs at TU Berlin in 2011 compared to previously published results from various groups
Fig.2. A successful VCSEL research effort must combine expertise in material growth and device fabrication expertise with dedicated characterizati on capabilities, such as measurement robots that can candle thousands of devices
Steps that we took to hit these record-breaking speeds with our 980 nm VCSELs included a reduction in the cavity length to half the wavelength of light, and a decrease in the reflectivity of the out-coupling mirror to cut photon lifetime and realize higher modulation speeds. We used oxide aperture layers made from 20 nm-thick layers of Al0.98Ga0.02 As that were positioned very close to the active region to reduce parasitic capacitance. These apertures were positioned very close to the active region to avoid damping of the resonance frequency by carrier transport.
We produced a range of 980 nm VCSELs with oxideaperture diameters from 1 μm to 10 μm. Designs had six different active media, between three and seven quantum wells with widths of 4 nm to 6 nm, and barriers with and without phosphorous. Unlike our previous generation of VCSELs, these lasers had an abrupt interface, rather than a graded one, around the active region. This change improved carrier confinement at high temperatures.
Testing this portfolio of devices revealed that VCSELs with aperture diameters between 5 μm and 7 μm could combine high output powers with high modulation bandwidths. Typical characteristics for a 6 μm diameter oxide aperture VCSEL were a threshold current of 0.9 mA at 20 °C, a peak power of 8 mW at a rollover current of 22 mA, and continuous wave operation up to 200 °C.
Fig 2. The gain in energy-efficiency resulting from the transition from copper to optical interconnects are substantial. However, a VCSEL is another two orders of magnitudes more energy efficient that optical interconnects derived from long-haul data-com modules. The VCSELs demonstrated recently in Berlin gain yet another order of magnitude
VCSELs with a 5 μm aperture delivered the most impressive performance at low temperatures. Buttcoupled to a 3 m multi-mode fiber, these devices enabled data transmission rates of 49 Gbit/s at -14 °C and 47 Gbit/s at 0 °C. Larger apertures were needed at higher temperatures to deliver more power to the detector. Transmission at rates of 12.5 Gbit/s, 17 Gbit/s and 25 Gbit/s were possible at 155 °C, 145 °C and 120 °C, respectively. These results illustrate that a laser – especially a VCSEL – is a highly nonlinear, coupled system. Gains in operating temperature come at the expense of transmission speeds, and VCSEL performance can be improved substantially with chip cooling.
Efforts are already underway at bringing these advanced VCSEL devices to market. TU-Berlin´s spinoff company, VI-Systems GmbH is working on providing detectors and devices for the commercial 850-nm waveband, and we are working with them to develop a VCSEL operating in this spectral range. We have found that the devices, which are closer to those serving the existing multi-mode VCSEL market than our 980 nm designs, require just one-tenth of the energy of a regular VCSEL to send a bit (see Table 1 for details).
Table 1. A comparison of the best results in energy-efficient data transmission, including values for the energy-to-data-ratio (EDR) and the heat-to-bit rate-ratio (HBR). The recently achieved results in Berlin won the Green Photonics Award 2012 at Photonics West in San Francisco
The energy consumption by these VCSELs is so frugal that they can meet the projections of the International Technology Roadmap for Semiconductors for energy efficient interconnects of 2015. Hitting this benchmark today proves that VCSELs are a technology that can already deliver the low-cost, energy-efficient, high-speed interconnects needed to overcome the bottleneck brought about by copper.
Future requirements on VCSELs
Even if these novel devices can accommodate the demand for the next few years, it would be naive to believe that no further progress is needed. Fortunately, there is the potential for far higher speeds from monolithic electro-optical modulated VCSELs – according to our studies, they could hit serial bandwidths of around 100 Gbit/s. And advanced coding schemes should lead to further gains in serial bandwidths and link robustness.
Another trend that we can expect to see is the development of high-contrast meta-structures, which should open up degrees of freedom in device design, such as multiple wavelengths on a single VCSEL array. Meanwhile, if there are moves to longer wavelengths, like 1.3-1.6 μm, this will enable compatibility with silicon waveguides that form part of optical silicon-based chips.
The spectral range is accessible with different classes of laser: InP-based VCSELs, which are well understood and commercially available from the likes of Vertilas and Beam Express and quantum-dot devices, grown on cheaper GaAs substrates.
Looking further into the future, when data traffic hits levels that are unthinkable today, even the footprint of a VCSEL might be too big. To address this, we are working in partnership with researchers at the University of Illinois at Urbana Champaign to develop nano-cavity lasers with metal cladding. This will shrink the output of optical sources for interconnects by a further two orders of magnitude, taking us into a realm where interconnects based on copper are never, ever considered.