Optical Interconnects for High-Performance Computing Systems

Optical interconnects employ light to transmit data between components of high-performance computing (HPC) systems, offering orders of magnitude greater bandwidth and lower latency than conventional electrical links. By integrating photonic devices onto silicon substrates, researchers have created compact, energy-efficient modules capable of handling multiterabit per second data flows. Key enabling technologies include photonic integrated circuits that combine lasers, modulators and detectors on a single chip, as well as wavelength-division multiplexing to multiply channel capacity. At the board and backplane level, novel optical fibres, waveguides and couplers reduce insertion loss and crosstalk, while on-chip photonic networks promise to overcome the bandwidth bottleneck among multicore processors. Advances in optical switching and mesh-based photonic network-on-chip architectures are addressing scalability and fault tolerance, and emerging modulators based on plasmonics or microresonators are pushing speed and energy-efficiency to new limits. Challenges remain in thermal management, packaging and integration with existing electronic control layers, but progress in hybrid electronic-photonic design is driving optical interconnects towards practical deployment in data centres, supercomputers and specialised AI accelerators worldwide.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent work has demonstrated ultra-low-crosstalk silicon Mach-Zehnder interferometer switches driven both thermally and electrically, achieving crosstalk suppression below –40 dB and switching times down to 100 ns. These devices leverage self-heating effects to balance phase shifts in each arm, minimising insertion loss and enabling high-fidelity 50 Gb/s signal transmission with on-chip losses under 10 dB. In parallel, a comprehensive roadmap of optical communications has highlighted the maturation of integrated photonic technology for datacentre and HPC backplanes, identifying trends in heterogeneous integration, co-packaged optics with electronic switches and the convergence of coherent optics with silicon photonics to meet the ever-growing demand for bandwidth and energy efficiency. At the on-chip scale, pioneering hybrid optical-electronic network-on-chip architectures have been proposed that combine an electronic mesh for short-distance traffic with bus-based optical channels for long-distance links. These systems employ dynamic power management to switch optical channels on and off in response to load, yielding energy-efficiency gains of up to 47 % compared with purely electronic networks and offering a pathway towards scalable, many-core processor interconnects.

Optical Interconnects for High-Performance Computing Systems publication trend

The graph below shows the total number of articles in optical interconnects for high-performance computing systems across all publications each year (not limited to Nature Index journals).

Technical terms

Optical interconnect: Data-transmission link using light rather than electrical signals, providing high bandwidth and low latency.

Silicon photonics: Integration of optical devices (lasers, modulators, detectors) on silicon wafers for mass-manufacturable photonic circuits.

Photonic network-on-chip: On-chip communication fabric that uses photonic links instead of metal wires to connect processor cores.

Mach-Zehnder Interferometer: Optical structure that splits and recombines light paths to control phase and amplitude for switching or modulation.

Crosstalk: Unwanted coupling of signals between adjacent channels or waveguides that degrades signal integrity.

References

1. Ultra-low-crosstalk silicon switches driven thermally and electrically. Microsystems & Nanoengineering (2025).
2. Roadmap of optical communications. Journal of Optics (2016).
3. Energy-efficient Hybrid Optical-Electronic Network-on-Chip for Future Many-core Processors. Elektronika ir Elektrotechnika (2014).