The high optical loss of typical photonic components, coupled with the low efficiency ofWDM-compatible lasers and the power consumed for microring trimming, dramatically increase the power consumption of silicon-photonic optical interconnects. Unfortunately, the majority of this power is typically wasted. While the full laser power is required to support periods of high interconnect activity, most of it is wasted when activity is low because the laser sources are always on, even when the interconnect transmits no messages. In addition, the power used to tune the network's microrings is affected more by the thermal variations imposed by processor activity, rather than interconnect utilization. Thus, the power drawn by a photonic interconnect is not proportional to interconnect activity. This chapter reviews some of the recently proposed techniques to achieve energy proportionality in optical interconnects, and provides directions for future research. In particular, Section 11.1 discusses techniques to turn off the laser source (i.e., "power-gate" the laser) when the network is idle in order to minimize the laser's energy consumption. To avoid exposing the laser turn-on delay, these techniques employ mechanisms that predict when communication is imminent and turn on the laser just in time for the transmission. Section 11.2 discusses a technique to minimize the energy consumed to keep the microring resonators within a narrow temperature zone. The technique consists of encapsulating the photonics layer in a porous silicon insulator to keep the thermal energy of the microring heaters from escaping the photonics layer, and employing microfluidics to minimize the impact of high thermal variations from the logic die.
|Original language||English (US)|
|Title of host publication||Photonic Interconnects for Computing Systems|
|Subtitle of host publication||Understanding and Pushing Design Challenges English|
|Number of pages||33|
|State||Published - Jun 30 2017|
ASJC Scopus subject areas
- Computer Science(all)