http://static.googleusercontent.com/media/research.google.com/fr//pubs/archive/42180.pdf

Interesting read, stated differently, but a solution is required

Abstracts :

Innovations in photonic integration circuit (PIC) and optical packaging techniques are necessary to maintain the scalability of next generation datacenter systems. Optics has evolved from long distance communication to short reach interconnects linking servers to switches, and switches to switches in modern datacenters. As the speed of datacenter switching systems increase, signal speed will keep increasing, leading to even more challenges in integration and packaging - higher speed means higher baud rate and/or more signal lanes, as well as higher power consumptions. Optical interconnects outperform their electrical counterparts at longer distances for higher speed signals [7], and will become more & more ubiquitous as data rates increase. However, compared to electronic design and manufacturing, optics is still lagging far behind in integration and automation. The pent-up demand in bandwidth will drive innovations on integrated photonics and packaging designs. In the long run, photonic integration circuits hold the key to enhance system functionality and reduce the size and power consumptions of optical transceivers, from short reach interconnects to long-haul coherent transmission [8].

Ok this too :

Modern mega-data centers employee a large number of 10Gb/s interconnects from servers to switches and in between switches [5], for a distance from 2m (servers to switch connection) to 2km (for switch connections between buildings). At such speed and distance, it is difficult for copper connections to meet the performance requirements. As datacenter capacities increase, both the speed and number of interconnects need to increase. A scale-out datacenter fabric typically involves thousands of interconnect links [6]. For the same bisectional network bandwidth, the efficiency and cost of the fabric strongly depend on the port count on each switch chassis, which is in turn limited by the bandwidth of each switch silicon and front panel transceiver density. To achieve the best front panel I/O density, low-cost, low-power and low-profile optical transceivers are thus of utmost importance in scaling out datacenter infrastructures. In order to maintain low cost and low power, it is important to not overdesign the performance of optical interconnects used inside datacenters. For short reach interconnects up to a few hundred meters at 10Gb/s, VCSEL based multi-mode optical transceivers are very low-cost and consume very low power. To achieve the 2km transmission distance requirements, single mode transceivers are preferred. Moreover, because of the large number of interconnects and much longer reaches involved in a scale-out cluster fabric implementation, the cost of optical fiber itself is significant in modern mega-scale datacenters. The cost of single mode fiber is intrinsically much lower than that of multi-mode fibers. So single-mode transceivers are becoming a trend for datacenter interconnect to not only save the cost of fibers but also improve the cabling efficiency and provide much longer reaches and futureproofing inside datacenters