Switch fabrics aren't new to multiprocessing. In the last decade, high-end MP systems, especially those for the military or those performing wide multimedia processing, reached the point that existing bus infrastructure couldn't support MP processing. There simply wasn't enough bandwidth in the existing bus structures, such as the VMEbus. So designers added adjunct buses, or mechanisms to increase the available bandwidth. These included speedups on VME (VME320), FPDP, and two parallel switch-fabric buses.
Mercury Computer and Sky Computer each developed its own parallel switch fabric, which switched parallel bus connections carrying data packets between the nodes. Mercury's RACEway and Sky's SKYchannel both became standards. They connected processor subsystems on a board, between boards, and even between boxes. However, their parallel connections, 32 bits for RACEway and 64 bits for SKYchannel, simply had too large a connection overhead. They weren't economical for very large systems.
But in the 1990s two new technologies emerged. These were high-speed serial buses and crossbar switches. Serial buses were implemented in communications and peripheral connections, while the crossbar provided the base for network and Ethernet switches. Combining those two—fast serial buses and crossbar switches—created the new technology of serialized switch fabrics.
This new technology had the low-pinouts of serial buses, plus the dynamic switchability and increased bandwidth of crossbars. (A crossbar can have n/2 connections for an n-port device. This is an improvement of n/2 in bandwidth over a parallel bus that has only one transfer at a time.)
The reason why the technology moved from parallel to serialized switch fabrics is easy to see. The latest version of RACEway runs at 66 MHz, delivering a peak bandwidth of 266 Mbytes/s. In contrast, the initial version of the RapidIO switch fabric will deliver 1.125 Gbytes/s of bandwidth for a 16-bit pseudoserial interface.
