High bit rate transmission must be matched by switching capacity. But switches are a big spender concerning energy and latency. This can be better! Click image to read more..

Today all switching is done in Store-and-forward Switches. A packet has a size between 72 up to appr. 2000 Bytes for an extended packet. Every packet has first to be stored and then forwarded to the other connection after the destination address has been read and understood. Since the switch makes forwarding decisions amongst others based on the destination address which is at the header of the packet, the switch can make the forwarding decision before receiving the complete packet, this process is called cut-through, the switch forwards part of the packet before receiving the complete packet. Cut-through allows lower latency and saves buffer space, but if an error occurred in the packet while utilizing cut-through, the packet will be forwarded with an error, alternatively, utilizing store-and-forward allows the switch to drop erroneous packets. For Cut-through, the incoming line and the outgoing line must have the same speed. If the speeds are different, normal store-and-forward technology has to be applied.

A state-of-the-art switch has a cut-through latency of 300 nanoseconds. In case of e.g. a switch with 10 Gbps servers and 40 Gbps interconnect lanes, the average delay per packet (no cut-through) will be 1000 nsec for store-and-forward, and 300 nanoseconds for the processing, totaling to 1.3 μsec.

Not only the latency is substantial, performing these tasks of identifying the address, store-and-forward the information, translating from optical to electrical and back again, costs a lot of energy and requires cooling.