Industry News
Phase Change and Silicon Photonics Accelerate AI
Views : 157
Author : JIUZHOU
Update time : 2026-10-12 09:36:43
As the scale of large-model training and inference continues to expand. Data exchange within AI clusters has become increasingly frequent. Traditional data center networks are primarily built on packet-switched architecture.
They are well-suited for handling Internet traffic with dispersed sources and frequently changing paths. Communication within AI clusters, however, exhibits characteristics distinct from those of traditional data center networks.
During specific training phases or for specific tasks, sustained, high-bandwidth data exchange requirements arise between certain nodes. These characteristics distinguish AI cluster communication from traditional data center networks. How can these requirements be identified and addressed using more appropriate connection methods? This is a challenge that intelligent computing infrastructure must resolve.
However, communication within AI clusters exhibits distinct characteristics. During specific training phases or for specific tasks, sustained, high-bandwidth data exchange demands arise between certain nodes. Identifying these demands and supporting them with better connection methods is now a key challenge for AI computing infrastructure.
Against this backdrop, optical switching technology has gradually come into focus. Optical Circuit Switching (OCS) establishes direct connections between nodes by creating end-to-end optical paths. Compared to traditional electrical switching, OCS reduces intermediate forwarding and processing steps and can adjust connection configurations based on service requirements.
In AI clusters, optical switching supports fast, high-bandwidth links between nodes. It also adapts to different tasks by reconfiguring optical paths. It does not replace all existing electrical switching equipment in the network.
Instead, it works with electrical switching and control systems. Together, they form a new network architecture for intelligent computing scenarios.
Currently, the OCS technology roadmap primarily includes two categories: spatial optics and guided-wave optics. Spatial optics solutions often use small mechanical parts to change how light travels.
Guided-wave optics routes light through chips, it is small, highly integrated, and easy to mass-produce. Silicon photonics is a key direction.
In silicon photonics optical switches, thermo-optic and electro-optic modulation are relatively common technical approaches. Thermo-optic modulation changes the refractive index by changing the local temperature.
The technology is fairly mature. Electro-optic modulation uses an electric field to control the material’s optical response. For large-scale integrated applications, both approaches require ongoing optimization of power consumption, thermal management, drive conditions, and device consistency.
In recent years, phase-change materials have opened up a new path for the implementation of silicon photonic OCS. Phase-change materials show different optical properties in different states.
By setting a state with an electrical signal, they can change how light passes through the waveguide. This enables the setup and switching of optical paths. Unlike continuous tuning methods, phase-change materials keep their written state without needing continuous power.
This non-volatile characteristic allows optical paths to retain their set state without a continuous power supply after programming. Thereby reducing the system’s static power consumption; when network topology adjustments are required. The paths can be reprogrammed via pulses to achieve nanosecond-level optical path reconfiguration.
Furthermore, combined with the high integration and CMOS back-end process compatibility of silicon photonics platforms. Phase-change OCS features a compact size, low static power consumption, and rapid reconfigurability. Making it suitable for high-density optical switching applications in AI computing networks.
They are well-suited for handling Internet traffic with dispersed sources and frequently changing paths. Communication within AI clusters, however, exhibits characteristics distinct from those of traditional data center networks.
During specific training phases or for specific tasks, sustained, high-bandwidth data exchange requirements arise between certain nodes. These characteristics distinguish AI cluster communication from traditional data center networks. How can these requirements be identified and addressed using more appropriate connection methods? This is a challenge that intelligent computing infrastructure must resolve.
However, communication within AI clusters exhibits distinct characteristics. During specific training phases or for specific tasks, sustained, high-bandwidth data exchange demands arise between certain nodes. Identifying these demands and supporting them with better connection methods is now a key challenge for AI computing infrastructure.
Against this backdrop, optical switching technology has gradually come into focus. Optical Circuit Switching (OCS) establishes direct connections between nodes by creating end-to-end optical paths. Compared to traditional electrical switching, OCS reduces intermediate forwarding and processing steps and can adjust connection configurations based on service requirements.
In AI clusters, optical switching supports fast, high-bandwidth links between nodes. It also adapts to different tasks by reconfiguring optical paths. It does not replace all existing electrical switching equipment in the network.
Instead, it works with electrical switching and control systems. Together, they form a new network architecture for intelligent computing scenarios.
Currently, the OCS technology roadmap primarily includes two categories: spatial optics and guided-wave optics. Spatial optics solutions often use small mechanical parts to change how light travels.
Guided-wave optics routes light through chips, it is small, highly integrated, and easy to mass-produce. Silicon photonics is a key direction.
In silicon photonics optical switches, thermo-optic and electro-optic modulation are relatively common technical approaches. Thermo-optic modulation changes the refractive index by changing the local temperature.
The technology is fairly mature. Electro-optic modulation uses an electric field to control the material’s optical response. For large-scale integrated applications, both approaches require ongoing optimization of power consumption, thermal management, drive conditions, and device consistency.
In recent years, phase-change materials have opened up a new path for the implementation of silicon photonic OCS. Phase-change materials show different optical properties in different states.
By setting a state with an electrical signal, they can change how light passes through the waveguide. This enables the setup and switching of optical paths. Unlike continuous tuning methods, phase-change materials keep their written state without needing continuous power.
This non-volatile characteristic allows optical paths to retain their set state without a continuous power supply after programming. Thereby reducing the system’s static power consumption; when network topology adjustments are required. The paths can be reprogrammed via pulses to achieve nanosecond-level optical path reconfiguration.
Furthermore, combined with the high integration and CMOS back-end process compatibility of silicon photonics platforms. Phase-change OCS features a compact size, low static power consumption, and rapid reconfigurability. Making it suitable for high-density optical switching applications in AI computing networks.
Related News
Read More >>
AI Data Center Optical Evolution: Trends & Market
Sep .20.2026
With the rapid growth of artificial intelligence computing, data centers now need much more bandwidth. The physical limitations of traditional copper cable connections are becoming increasingly apparent as transmission speeds rise.
The Global OCS Market Will Exceed $8 Billion by 2030
Sep .09.2026
The latest industry forecasts indicate that by 2030, the total size of the global optical path switch market is expected to exceed $8 billion, with a compound annual growth rate (CAGR) of over 30% between 2026 and 2030.
400G Single-Lane: Next Inflection for AI Datacenters
Aug .27.2026
400G per lane is the next inflection for AI data centers, boosting bandwidth and scaling, but requires advances in materials, packaging, and cooling.
AI-driven 800G Ramp-up Amplifies Quality Risks
Aug .14.2026
Faced with a surge in orders, leading global manufacturers are increasing production line speeds and shortening delivery cycles. However, the real challenge of this growth goes beyond “can we deliver orders on time?”