Optical technologies offer the potential for significantly increased performance with lower power than semiconductor electronics at least for certain elements of a PetaFLOPS system. This is a consequence of the fundamental differences in the device physics of the two approaches. Unlike electrons, photons do not interact as they cross paths, resulting in a number of desirable properties. Although not as mature as semiconductor technology, optics are already having an impact on medium- to long-range communications and high-density secondary storage. Both these areas have the potential for significant growth leading to critical contributions for PetaFLOPS computing. One area for which optics does not appear to be well suited is in the implementation of logic gates. Thus, while optical technology may be essential for certain critical components, it is most likely to be incorporated in a hybrid architecture integrating two or more distinct technologies.
Optical communications technology for computer module interconnect has emphasized medium- to long-distance paths where its performance benefits over conventional wire-based media offset its current cost disadvantages. As bandwidth requirements increase, optical methods will become favorable for short distances, perhaps even for chip-to-chip interconnect. Optical communication methods exhibit higher bandwidth capacity by orders-of-magnitude than electrical means and at sufficiently high data rates impose substantially lower energy penalties. These high bandwidth and efficiency advantages are reinforced by optical technology's electrical isolation properties, greatly reducing the possibility of cross-coupling which would otherwise degrade reliability.
There are two primary forms of optical interconnect: guided wave and free-space. Guided wave optical communication employs fiber optics or wave guides to direct light signals between two fixed points. Where line-of-sight paths exist, free-space systems permit high-density space multiplexed signal packing and the potential for path switching and one-to-many broadcast communication.
The state of the art in guided wave optical communication technology provides 100 megabits per second (Mbps) using light emitting diodes. Recent advances in laser diode emitters has achieved bandwidths of 2.5 gigabits per second (Gbps). Free space optical interconnects using symmetric, self-electrooptic effect devices have shown the capability of 150 Mbps. High-density, thousand-channel, free-space ``fabric'' systems have been developed producing throughputs of up to 150 Gbps. Arrays of laser diodes have been fabricated on single semiconductor dies demonstrating the feasibility of electrooptic interfaces and high-bandwidth inter-chip optical communication.
It is projected that within 20 years when PetaFLOPS systems will be feasible, guided wave technology will be capable of providing throughputs on the order of a million million bits per second or 1 terabit per second (Tbps). Using vertical-cavity surface emitting laser diodes, free space methods may reach a capability of 10 million billion bits per second or 10 petabits per second (Pbps). Not only are these levels of throughput necessary to support PetaFLOPS scale computation, but the added advantage of free space interconnect not requiring the potentially millions of point-to-point wire/fibers to be connected may prove essential for reliability and economic manufacture.
The other area in which optics is expected to have a major impact on PetaFLOPS system design is in the area of memory and mass storage. These take the form of planar and 3D technologies. CD-ROMs and optical tape are the two most common examples of planar optical storage. The consumer level CD optical storage holds somewhat less than 1 gigabyte and industrial scale optical disks have capacities of up to 20 gigabytes. Optical tape systems have capacities of between 50 gigabytes and 1 terabyte. Access is slow with access times measured in milliseconds for on-line disks and many seconds for tapes and robot-loaded optical disks.
Still at the research stage, 3D optical storage techniques offer the prospect of extroadinary memory capacity and bandwidth at moderate to high access speeds. Techniques such as photorefractive rods, two-photon 3D memory, and spectral hole burning are being pursued in the laboratory. It is projected that within ten years 2-photon holographic techniques with spectral hole burning and acousto-optic scanners will provide memory systems on the scale of 10 terabytes with bandwidths of one Pbps and access time of a microsecond. Using 2D spatial light modulation, storage capacity of 10 petabytes may be achievable in 20 years.