Imec Debuts 100 GHz Photodiode to Enable 400 Gbps Optical Links for AI Data Centers

At the European Conference on Optical Communications in September 2026, Belgian semiconductor research center Imec demonstrated a 100 GHz germanium-on-silicon avalanche photodiode operating at just 5 volts. The component enabled the first net 400 Gbps optical link using an avalanche photodiode receiver, establishing a new hardware baseline for the massive data exchanges required inside artificial intelligence data centers. By proving that high-speed optical reception can be achieved without prohibitive power demands, the demonstration addresses one of the most pressing physical bottlenecks in modern hyperscale computing environments.[1][4]

As artificial intelligence training clusters scale to accommodate increasingly complex models, the physical infrastructure connecting individual server racks faces a strict bandwidth-to-power bottleneck. Facilities must move terabits of data per second with minimal latency, pushing conventional optical receivers to their absolute physical limits. Avalanche photodiodes inherently amplify incoming optical signals to improve receiver sensitivity, making them a highly desirable solution. Historically, however, engineers had to sacrifice overall bandwidth or accept impractically high operating voltages to achieve that built-in signal gain, limiting their deployment in high-speed data center interconnects.[2][4]

To break that historical trade-off, Imec researchers fundamentally altered the standard separate absorption, charge, and multiplication architecture that governs how these photodiodes function. The engineering team scaled the multiplication layer to less than 100 nanometers by leveraging a deeply recessed germanium-in-silicon structure, while completely removing the traditional charge layer. This structural revision allowed the semiconductor to process optical signals at extreme speeds without requiring the high voltage typically needed to drive the avalanche multiplication process, effectively rewriting the operational parameters for silicon photonics receivers.[1][4]

That precise structural manipulation allowed the device to maintain a 100 GHz bandwidth and a responsivity of 1.8 A/W across both the O-band and C-band spectrums while drawing only 5 volts of power. The resulting hardware delivers a 3 decibel improvement in receiver sensitivity, which translates to approximately a twofold internal gain over standard, non-amplifying photodiodes. Achieving this level of performance across multiple optical bands provides system architects with the flexibility needed to design diverse interconnect layouts without swapping out the underlying receiver hardware.[1][2][4]

"The APD provides additional link margin compared with a conventional photodiode," said Joris Van Campenhout, vice president of research and development for optical interconnects at Imec, during the conference presentation. "Combined with its high bandwidth and low operating voltage, this makes the device particularly attractive for future scale-up optical interconnects." The ability to maintain that margin at 5 volts is the critical factor that elevates the component from a laboratory curiosity to a viable candidate for commercial data center deployment.[1][4]

For system architects designing the next generation of artificial intelligence clusters, that 3 decibel margin serves as a highly flexible currency in power-constrained environments. Facility operators can spend the margin to reduce the power consumption of the transmitting lasers, cutting overall facility energy use and cooling requirements. Alternatively, they can use the improved sensitivity to tolerate higher optical losses in complex, densely packed server racks where cables must navigate tight physical spaces and multiple connection points without degrading the data stream.[1][3]

To prove the system-level viability of the new receiver, Imec paired the photodiode with a beyond-110 GHz electro-absorption modulator that the organization had introduced at the 2025 iteration of the same conference. Operating together on Imec's standard 300-millimeter silicon photonics platform, the paired devices successfully maintained the 400 Gbps optical link. By utilizing an established 300-millimeter manufacturing platform, the research hub demonstrated that the breakthrough does not require exotic fabrication techniques, clearing a major hurdle for eventual mass production by commercial semiconductor foundries.[1][4]

With the fundamental architecture proven, the research group is now shifting its focus toward commercial hardening and integration. Subsequent engineering phases will test the photodiode's reliability across varying temperature ranges and fluctuating optical input power conditions, ensuring it can survive the harsh thermal realities of an active server rack. Simultaneously, the team will work to integrate the receiver directly with high-speed electronic components on the 300-millimeter production line, moving the technology one step closer to deployment in the global artificial intelligence supply chain.[1]