, Columbia University
, Hewlett-Packard Laboratories
Pages: pp. 3-5
Abstract—This introduction to the special issue summarizes the creation of the first Optical Interconnects Conference, discusses some key themes in the field, and introduces the seven articles included in the special issue.
Keywords—optical interconnects, high-performance computing, data centers, datacomm, silicon photonics, wavelength division multiplexing
Computational power at all scales, from personal devices to supercomputers and data centers, continues to grow. But, over the past decade that growth usually hasn't resulted from performance increases in single microprocessors. Instead, it has come from increasing numbers of microprocessors, both on a single chip and in the complete system. Those processors need to communicate between each other and with the associated DRAM. It is by lashing the processors together through the interconnect network that supercomputers have continued to increase their performance and data centers have grown ever larger.
Over this same period of time, our long-haul, metro, and even campus-level communications backbones have largely been converted to optical communications. Even cable television companies now utilize optics for most of their links, and in some areas an optical fiber is connected directly to our homes. The dominance of optics for gigabit-per-second communication links longer than a few hundred meters has been driven by electrical links' inability to achieve the required performance.
These two trends—dramatically increased computational power and the growing use of optical links—are now being combined as the optics research community tries to use new optical communications technologies to improve interconnects between, and perhaps someday within, the chips that power supercomputers and data centers. Recent well-attended sessions related to optical interconnects at the Optical Fiber Communication Conference (OFC) and other large conferences have highlighted the recent progress and tremendous interest in this area by researchers. Thus, there was a need for a conference focusing on the subject, and so the IEEE Photonics Society decided to sponsor the first Optical Interconnects Conference. For 22 years, the Photonics Society had sponsored a small workshop on interconnects within high-speed digital systems in Santa Fe, New Mexico. With the dedicated help of that workshop's existing program committee, we were able to put together the first Optical Interconnects Conference in just a few months. The conference was held in Santa Fe on 20–23 May 2012. With 50 oral papers (including invited and plenary) and 29 poster papers, the 177 conference attendees were able to see the emerging ideas in the field. This special issue of IEEE Micro explores these ideas with seven articles that we selected in our role as conference general chairs. We greatly appreciate the assistance of the program committee chairs, Ashok Krishnamoorthy, David Miller, and John Shalf, in the selection process. The second Optical Interconnects Conference is scheduled for 5–8 May 2013.
Research in optical interconnect hardware today follows two distinct directions. The first utilizes parallel multimode optical fibers and is based on III-V semiconductor devices, especially photodiodes and vertical-cavity surface-emitting lasers (VCSELs), which are used in most datacomm optical links today. Work in this area is aimed exclusively at interconnects between silicon chips, and requires a short electrical link between the chip and the optical components. "High-Bandwidth Optical Interconnect Technologies for Next-Generation Server Systems" by Kazuhiro Tanaka and his colleagues follows this direction, showing how to assemble a low-power, low-cost, high-speed parallel VCSEL link that can essentially replace an electrical cable assembly while yielding higher performance and density at lower power. The authors also note that an optical link to system memory could have advantages. Past research in this arena has already led to successful implementation of parallel optical interconnects in many of today's highest-performing supercomputers as listed at http://www.top500.org.
In "Optical Interconnects for High-Performance Computing Systems," Michael R.T. Tan and his colleagues show how to push this VCSEL technology into a multicast optical backplane bus, which could also allow interconnects to move away from the point-to-point links that were required when electrical link bitrates exceeded approximately 100 Mbits per second (for example, the transition from conventional PCI to PCI Express). This move back to a true communications multicast bus would lead to a significant rearchitecting of multiprocessor systems.
The other major theme of optical interconnects today is to build a technology that is compatible with conventional silicon CMOS processing, which would ultimately allow integration of optical communications on a chip. This technology could therefore be used for communication between chips and also within a chip. Today, companies such as Luxtera have demonstrated that this silicon photonics (SiPh) technology is feasible by commercially manufacturing discrete parallel optical modules based on SiPh. One key issue is that silicon, as an indirect bandgap semiconductor, has not yielded a good light source by itself, so modules such as those provided by Luxtera use an external indium phosphide (InP) laser. "Hybrid Silicon Devices for Energy-Efficient Optical Transmitters" by Sudharsanan Srinivasan and his colleagues outlines attempts to solve this issue by oxide bonding direct-gap III-V semiconductors onto silicon. If successful, the approach also lends itself to including other III-V active devices such as modulators and detectors onto silicon.
Another issue with SiPh technology is that the fabrication of active and passive optical devices can only be done on silicon-on-oxide (SOI) wafers. Furthermore, the SOI wafers used for SiPh have a far thicker layer of silicon than those of conventional SOI electronic chips. In "Monolithic Silicon Waveguides in Standard Silicon," Chia-Ming Chang and Olav Solgaard point to a potential solution with an initial demonstration of passive optical waveguides fabricated on bulk silicon wafers. The next step for this work would be to demonstrate active devices compatible with these waveguides.
For SiPh to become a truly valuable high-bandwidth technology, multiple wavelengths of light must propagate in the same optical waveguide or fiber, with the wavelengths combined or separated at the transmitters and receivers.
This wavelength division multiplexing is a standard technology in long-haul optical telecommunications, but maintaining channel separation typically requires temperature stabilization of the optical components to approximately 1°C, which could be quite difficult in a chip environment. In "Bit-Error-Rate Monitoring for Active Wavelength Control of Resonant Modulators," William A. Zortman and his colleagues take a step toward solving this issue by showing how to set up a feedback loop to lock the center wavelength of a SiPh optical modulator on that of a laser channel.
These three SiPh articles are concerned with developing SiPh device technologies. If SiPh is realized as a technology, it can allow substantial changes in chip and system architectures. In "Silicon Photonic Microring Links for High-Bandwidth-Density, Low-Power Chip I/O," Noam Ophir and his colleagues propose such an architecture, make some assumptions about the required SiPh devices, and examine the system-level performance that would be achieved. Finally, "Silicon Photonic Interconnects for Large-Scale Computer Systems" by Ron Ho and his colleagues radically re-envisions the architecture of a silicon chip with integrated SiPh. The authors propose the concept of a "macrochip" composed of small chips interconnected by a SiPh optical network. The authors go on to show their initial work aimed at achieving this goal.
Much more progress is required in optical interconnects—at the device and link levels and in the architecture—to bring the technology to fruition. But if the technology is successfully implemented, it has the possibility of substantially enhancing the performance of future computing systems. The history of electronic computation has shown an exponential performance increase with time; optical interconnects can allow this trend to continue into the future.