A crucial component in a future teraflop, distributed, massively-parallel processing system is an interconnection network capable of reaching an aggregate capacity of a terabyte-per-second with gigabyte-per-second access rates for each user. This is the target of the architecture described here, which is the optoelectronic extension of the interconnect previously used in a shared-memory, distributed, MIMD supercomputer. Because of its inherent flexibility, at a lower performance level this same architecture could also meet the growing near-term interconnect needs within and between current large computing centers. The architecture provides small block transfers, a source-destination latency approaching the physical minimum, high average and burst user bandwidth, and expandability to thousands of distributed users. The packet payloads remain in optical format from source to destination while being switched in the multi-state, multi-path network. Large-scale data compression by wavelength results in packets of a few nanoseconds length. Self-routing at intermediate nodes is achieved with brief excursions of the control headers into the electronic domain, where pipelined processors implement a hot potato/deflection switching protocol. A system demonstration is currently under construction.