R&DRecommendations

A considerable amount of research and development remains to be done to reach the projected performance levels listed above for the optics technologies. Listed in Table 5.9 are the recommendations of the committee on research directions that need to be pursued in support of any national thrust to develop PetaFLOPS computing systems. The dollar amounts associated with each recommendation are rough estimates of the investment needed over the next 10 years to ensure timely insertion of the technologies. The first two items, smart pixel arrays and interconnection optics, are already receiving some support, but the recommendation is to expand the support in these areas. The remaining entries in Table 5.9 are recommended research programs to support the insertion of optics technology into PetaFLOPS computing systems.

A brief discussion of each of the recommended research areas follows:

Smart Pixel Arrays

provide the interface between parallel electronics and parallel optics in that each element (pixel) of the array has both electrical and optical inputs and both electrical and optical outputs, and they provide a localized processing capability at each pixel. These arrays will be critical for the realization of large-scale parallelism for optical interconnection and optical memory addressing. Research is needed in integrating the electronics and optics, improving thermal management, and providing adequate interfaces to the arrays.

Interconnection Optics

includes conventional optics needed in the guided or free-space channels. Research is needed to miniaturize light shaping and directing elements (e.g., lenses and computer-generated holograms), to realize single-mode fiber arrays, and to create an optical ``HIPPI'' standard for 100-gigabytes/sec links.

Optical Memory Backplane

calls for the development of optical interconnections between memory boards for the purpose of realizing required fanouts in memory addressing. As mentioned previously, optical interconnection avoids the capacitive loading problems associated with fanouts in electrically interconnected systems that cause an increase in the RC time constant. In addition, the outputs of the memory chips have to be ORed together, creating quite a load on the drivers of these chips, thus slowing them down. Optical ORing would be faster.

3-D Memory Addressing

needs research to search for efficient ways to store and read out such large data stores, and to develop the required technologies for rapidly focusing the addressing beams to the desired spots within the 3-D storage media.

Interface to Josephson Junction CPUs

calls for research to develop optical links that can provide interconnections with the outside world for processors inside the cryogenic environment. The problem with electrical connections for this application is that they conduct heat into the cryogenic chambers.

Optical I/O for ICs

is intended to address the packaging issues related to providing ICs with optical interconnects, and to support a competition between the electronic and the optical communities to determine the demonstrated limits to how many I/O pins can be achieved on ICs by electrical and optical interconnects, respectively.

Optical CAD Tools

are the optical counterpart to the electronic CAD tools that have been so successful in enabling system designers to work with complex systems. Research is needed not only to develop needed tools, but to integrate all of the available tools into a user-friendly CAD system for optoelectronic design.

All-Optical Terahertz Logic

will be needed for high-speed controllers for switched optical links (e.g., the optical crossbar), for high-speed interface units needed to load and unload data to and from high-performance fiber networks, and to implement multiplexers and demultiplexers needed for TDM fiber networks.

Optical Radio

is an interesting concept that should be researched as a way of providing general interest data throughout PetaFLOPS systems (e.g., for distribution of instructions or maintenance of cache coherency). Many different optical signals are broadcast simultaneously to many different receivers which choose which signals to ``listen to'' at any particular time.