Optical on-chip data transmission enabled by silicon photonics (SiPs) is considered a promising candidate for future on-chip communication as the high-bandwidth, low-latency, and relatively-distance-independent nature of photonics overcomes the technological limitations of the electrical interconnects currently used in Chip Multiprocessors (CMPs). Present SiP technologies, however, impose static power overheads that often eliminate their performance and dynamic power benefits. Consequently, many research efforts have been focusing on both the technology and the architectural level to solve this static power problem and to unleash the full potential of SiPs. This thesis proposes and evaluates novel architectural approaches to allow for a more power-efficient utilisation of optical links in networks-on-chip (NoCs). First, it carries out a thorough review and analysis of the state-of-the-art optical NoC (ONoC) architectures and a comparison of current electrical and optical interconnect technologies in terms of latency and power consumption. Then, it proposes `Amon', a novel all-optical NoC based on wavelength-selective routing that decreases static optical power by exhibiting a topology with lower path losses and fewer wavelength switches. Moreover, Amon features an improved destination-reservation mechanism and backend modifications to further improve performance and power efficiency. Afterwards, it introduces `Lego', a novel hybrid NoC combining electrical and optical interconnects in an architecture in which high quantities of low-bandwidth optical links provide high bisection bandwidth with reduced power consumption due to lower overall optical losses. A distance-based routing mechanism ensures that optical links are used for large enough distances to hide their serialisation delay and an electrical NoC for local traffic otherwise. Finally, it presents a novel subchannel scheduling scheme and arbitration mechanisms for shared optical buses to improve bandwidth utilisation and power efficiency compared to the state of the art by scheduling contending nodes both in time slots and subchannels.