Optical fiber has an enormous capability for the transmission of digital information. One single-mode fiber can transmit several hundred Gigabits/sec, all within a core diameter of less than 5 microns. The challenge of optical networking is to harness this technology in a way that is well suited for internet-type (bursty) traffic in a cost-effective manner.

All-optical networks have been proposed as a partial solution. Wave-length-division multiplexing techniques and optical star couplers are ideally suited for constant bit-rate transmissions. Once a circuit has been established, there is an open channel for sending the data.

Asynchronous Transfer Mode (ATM) is a standard for high-bandwidth networking. The small cells minimize the latency. The fixed size of the cell facilitates an all-hardware switching engine. The requirements of ATM, however, suggest the use of an optical-electronic components, rather than all-optical systems, for the following reasons.

• Ordering: Adaptation Layer 5 (AAL5) does not include sequence numbers within the ATM cell. Therefore, the order of any sequence of cells generated at one source must be maintained.

• Header Translation: The ATM header (5 bytes) is modified as the cell is passed through the switch.

• Cell length: At a switching rate of 4 Gbps, a reconfiguration time of under 10ns is required in order to achieve a 90% utilization of the switch. Such reconfiguration rates prohibit the use of fiber Fabry-Perot elements, due to delays from mechanical inertia.

• Buffering: If multiple cells are to be transmitted to a single switch port simultaneously, all but one of the cells must be dropped, or deflected, or buffered.
  • If any cell is dropped, the entire AAL5 frame is lost, thus decreasing the true rate of useful transmission. Cell loss must be kept minimal.
  • If a cell within a stream is deflected, reordering can occur when the stream is recombined. This is not appropriate for AAL5.
  • Buffering, therefore, is required. Unfortunately, an optical random-access memory is not feasible. Silicon DRAM, on the other hand, provides fast access at low cost.

For the reasons listed above (header translation, buffering, and fast reconfiguration), we see ATM networking requires a mix of electronic and optical components. Thus, a Optical/Electronic conversion is needed at some points within the network.

Our microelectronics research investigates the integration of optical components on the same substrate as the transistors. Our systems research investigates the construction of an input-buffered ATM switch to maximize the throughput of the electronic memory buffers.