The rates at which computers can process information are beginning to outpace their abilities for them to transmit and receive it. Internet traffic will soon not only consist of millions of web pages, but also millions of on-demand video and audio streams destined for billions of digital media devices. But although processing power will advance steadily with new generations of multi-core processors, the ability to feed these chips with data is becoming limited. The electrical lines connecting chips, boards and systems will have difficulty keeping up with Moore's law.

Electrical is giving way to optical. Fiber has already taken over the long haul links used to carry telephone calls and internet traffic between and cities. Rings of optical fiber feed service providers with gigabits of optical data within major metropolitan areas. In data centers, 10 Gigabit Ethernet links are coming on line, relying on photonics to transport the data. The reason is clear -- while metal wires have difficulty moving gigabits of data further than a football field, individual fibers have the capability to carry terabits of data between cities.

Over the next decade, the copper links connecting chips and boards within servers and will also become strained, in speed, distance and power. For 20" chip-chip links, pushing line speeds above 10 Gbps requires sophisticated equalization techniques to compensate for distortion caused by imperfections in the copper and the underlying FR4 dialectic which comprises the motherboard. Each gigabit above 10 Gbps will become more expensive -- and even a major industry change such as a lower loss dielectric for the motherboard only buys a few gigabits of headroom. 20-gigabit may be possible -- but only if one limits link spans to six inches or less.

Figure 1 -- Simulation of  20"  channel transmitter w/ equalization.  
Switching the board material will only extend speeds by a small amount.

Rack and blade systems are crowded with copper lines which drain power, generate heat, and force system designers into placing the hottest components right next to each other. The result: some bladed servers can't be fully populated with the latest and greatest processors and workloads may have to be offloaded to cooler systems to avoid overheating. 

One solution... Optical, but the Challenge is Cost

The benefits of fiber optic links are well known. They have tremendous data capacity, a compact size, and the ability to carry data up to 80km or more without regeneration. Within a data center (distances up to about 100 meters) they can have near-zero attenuation. This provides distance independence that could free designers to re-architect systems in new ways which take into account performance, maintenance and thermal issues.  However, the main challenge just as well known -- optical components remain expensive for these data center based applications.  

These enterprise networking problems are motivating Intel's Photonics Technology Lab to turn to the semiconductor industry's bread and butter -- silicon. If optical functions could be manufactured and integrated in silicon wafers, just as electrical devices are today, the photonics industry could benefit from the billions of dollars of infrastructure and decades of R&D built up around silicon manufacturing over the past 40 years. By "siliconizing" photonics, the benefits of optical communication could be brought to future networks, servers, and PCs.

Silicon, to date, has hardly been considered a good material for optical components when compared to semiconductors such as Indium Phosphate and Gallium Arsenide. These materials can be made to readily emit, detect and transport light. However, after many years of development they still remain costly to manufacture, low in yield, and tricky to integrate different devices on the same chip. Devices are manufactured separately, and these discrete components must be assembled with complex, user-intensive active alignments in order to make sure that the laser beam travels from one component to the next. These limitations make silicon integration worthy of investigation.

The greatest promise of silicon photonics is the integration of different types of devices into a single silicon platform. Materials such as silica (glass) support passive devices to route light but lack active devices to manipulate light, while active materials such as indium phosphide require yield-killing regrowths to integrate different types of devices. Silicon, on the other hand, has both active and passive capabilities and the potential to integrate them with a much higher yield. If CMOS compatibility is maintained, photonic devices might even be monolithically integrated with electronics -- converging electronic computing with optical communication all on one single silicon chip.  

Six Building Blocks Are Required

Intel's research in silicon photonics is in the first of three phases. The first phase is to prove silicon's viability as an optical material. Silicon does have active and passive optical capabilities, but the active ones have been limited in performance. Through research, these capabilities must first be extended and expanded before enough functions exist to build useful integrated photonic modules. To this end, Intel has focused most of its efforts into active capabilities such as tuning, detecting, switching, modulating, and amplifying light. This research has produced a few recent success stories including the first gigabit speed silicon modulator and the first continuous-wave silicon laser.  

 Figure 2 Intel sees six major building blocks that must be addressed 
to make silicon photonic communication a reality.

Optical modulators are used to encode a high quality data signal onto an optical beam, effectively by turning the beam on and off rapidly to create 1s and 0s. Although 10Gbps modulators made from lithium niobate and indium phosphide are common today for long distance communication, before the year 2004 no one had built one from silicon that was faster than about 20MHz. In February of 2004, Intel announced the first gigahertz silicon optical modulator in the prestigious scientific journal Nature. By integrating a novel transistor-like device, Intel was able to create a modulator that scaled much faster than previous attempts. Just over a year later, Intel researchers further demonstrated that its silicon modulator was capable of transmitting data up to 10 Gb/s.

Silicon lasers had not been built in the past due to the material's semiconductor properties, which don't allow for the efficient emission of photons or even light amplification in the way materials such as Indium phosphide can. Instead, Intel researchers found that by using an effect called Raman scattering they could amplify light as it passes through silicon. This is accomplished by transferring energy from a second "pump" beam through the silicon crystal and into the passing signal beam.  However, the challenge was to maintain continuous operation -- due to a confounding quantum effect called "Two Photon Absorption," which caused a cloud of light-absorbing electrons to accumulate in the amplifier. By integrating a diode-like semiconductor device, the researchers found they could sweep out the electrons and achieve continuous operation. By surrounding this optical amplifier with mirrors (coated onto the ends of the chip), Intel then created the world's first continuous-wave silicon laser.

Once the silicon photonics research community has produced a suite of compelling active optical devices or building blocks such as those described above, Intel hopes to enter a second phase of hybrid integration. This will start with a silicon optical bench, in which electronics and photonics are passively aligned onto a micro-machined silicon assembly platform. Lithographically defined silicon structures guide components into the proper place without the need to even turn on the light. In this phase Intel would begin to integrate certain functions directly into the silicon. For instance, a silicon optical multiplexer and detectors could be integrated to form a multi-channel receiver. Or, silicon modulators, lasers, and passive components could be integrated to create a multi-channel transmitter. This will happen only when the integrated solution brings a benefit to the final module, which may be an increase in performance, a smaller form factor, or a lower total cost.  

Monolithic Silicon Integration

The final phase of research will be to approach monolithic silicon integration -- putting all of the components together to form modules that process light in ways that could never be done before and at costs low enough to allow them to become ubiquitous. This may include electronics if this brings a benefit and if photonic-electronic integration challenges can be addressed.

If the cost becomes low enough, new products will be possible. Imagine tiny, integrated silicon optical transceivers embedded directly into the connector of an optical cable, complete with an electrical interface. To a technician, this would look like simple electrical cable -- with all of the sensitive optical interfaces hidden inside the connector. Network equipment such as server blades could be built with electrical-only interfaces -- the conversion to optical would be transparently done within the cable itself. If the link fails, just replace the cable.

Integrated, high volume silicon photonic chips could dramatically change the way that enterprises use photonics links for their systems and networks. As already mentioned, simply having photonics could eliminate bandwidth and distance limitations, allowing for radically new flexible architectures capable of processing data more efficiently. Silicon photonics may even have application beyond digital communications, including optical debug of high speed data, expanding wireless networks by transporting analog RF signals, and enabling lower cost lasers for certain biomedical applications. 

As the market moves into a new era of multi and eventually many-core processors, the amount of data that computing platforms consume will continue to rise in accordance with Moore's Law. More and more, systems and networks built around these high-performance platforms will be limited by the data pipes connecting them to their neighbors. In addition, as users begin streaming content to billions of connected digital media devices, in the form of next generation PCs, cell phones, music/video players, HDTVs, and more, a crushing tide of data will hit tomorrow's networks. Intel's silicon photonics research may allow that tide to be easily and affordably channeled through hair thin optical fibers.

About the Author

About Intel's Photonics Technology Lab