The explosive growth of artificial intelligence is pushing data center networks beyond the limits of traditional electrical connections. Modern AI clusters contain hundreds or even thousands of GPUs that must exchange enormous volumes of data in real time. To keep those processors working efficiently, engineers are increasingly turning to optical technologies capable of delivering higher bandwidth and lower latency than copper wiring. A new photonic chip based on dense wavelength-division multiplexing (DWDM) could play a key role in that transition, tells IEEE Spectrum.
DWDM is an optical networking technique that transmits multiple light signals at different wavelengths through a single fiber. Instead of sending one stream of data through each fiber, multiplexing allows dozens of channels to share the same line, dramatically increasing capacity without adding more physical connections.
Researchers at Scintil Photonics, working with Tower Semiconductor, have developed a photonic integrated circuit called LEAF Light that integrates multiple DWDM laser channels directly onto a chip. The device incorporates arrays of distributed-feedback lasers along with photodiodes and modulators, allowing it to send several wavelengths through a single fiber connection. This design addresses one of the major bottlenecks in AI infrastructure: the communication links connecting processors inside dense computing clusters.
AI systems rely heavily on “scale-up” networking, in which accelerators within a rack or cluster must function almost like a single supercomputer. If the network cannot keep up with the processors, GPUs spend time waiting for data instead of performing calculations. Optical multiplexing can help solve this problem by increasing bandwidth while reducing latency and power consumption.
The LEAF Light chip sends eight or sixteen wavelengths through a single fiber, each channel carrying lower-speed data streams that combine into total speeds approaching 1.6 terabits per second. Distributing the load across multiple channels also improves energy efficiency and allows networks to maintain faster response times between processors.
Manufacturers expect the first production units to reach customers by the end of 2026, with wider deployment planned for large-scale AI systems later in the decade. If successful, DWDM-based photonic chips could become a core technology enabling the next generation of AI data centers.