Technical Insight
Giving AI room to grow with ultra-compact silicon modulators
Scaling AI infrastructure requires ultra-fast communications across distant processors, but conventional silicon modulators are bulky and power hungry. A new, ultra-compact silicon microring modulator has demonstrated record-breaking transmission speeds at low power, offering a solution that could give AI space to grow.
BY ALIREZA GERAVAND, LESLIE A. RUSCH AND WEI SHI, CENTRE FOR OPTICS, PHOTONICS AND LASERS (COPL), UNIVERSITÉ LAVAL, CANADA
AI is fundamentally transforming every sector, from healthcare and finance to automotive and consumer technology. This revolution is being powered by enormous computing clusters – thousands of computing nodes spread across multi-acre campuses—and as AI models grow in size, they require exponentially more processors working in parallel. OpenAI’s GPT-4, for instance, used 25,000 interconnected computing nodes over 90 days for training, but next-generation clusters are projected to involve 300,000 or even a million nodes in the near future.
To operate at this scale, massive volumes of data must be exchanged rapidly and efficiently, both between compute nodes within a cluster and between clusters. Optical interconnects form the backbone of these systems, providing high-speed, low-loss data transmission through optical fibres.
This relies on the efficient conversion of electrical data from processors into light—a process that is enabled by silicon photonics technology, which also benefits from mature manufacturing capabilities. However, the unprecedented communication demands between AI processors are creating a performance bottleneck – one that risks slowing the pace of AI advancement. Today’s optical links in AI clusters often rely on simple intensity modulation and detection – essentially turning a light on and off to send data. While cost-effective, this approach faces increasing challenges in meeting AI’s soaring requirements for both speed and reach.
Next-generation silicon photonics-based technologies must deliver far higher data capacity and longer reach, while scaling seamlessly as AI clusters expand. The solution is coherent optics, a technology that already dominates long-distance applications. Leveraging the coherent properties of light, we can go beyond the simple on/off flashlight approach and instead modulate both the amplitude and phase of light.
This advanced modulation dramatically increases transmission capacity, which explains its widespread adoption in long-haul networks. It also offers greater robustness to link impairments and can reach longer distances at the same signal power. However, conventional coherent optics has so far been bulky, power-hungry, and challenging to integrate with co-packaged optics (CPO). The primary culprit is the Mach-Zehnder modulator, the most commonly used modulator in coherent systems.
To achieve truly massive data rates, all available wavelengths must be exploited through wavelength-division multiplexing (WDM). While Mach-Zehnder modulators can handle
multiple wavelengths, they require many additional components, adding complexity, size, and cost. To overcome these limitations, we have developed a transmitter for coherent optical interconnects, featuring ultra-compact, energy-efficient coherent modulators integrated directly on silicon chips.
Rethinking coherent silicon optics
Central to our innovation is the microring modulator (MRM), a tiny structure etched onto silicon chips. Although MRMs are known for their compact size and low power
operation, they have historically been considered unsuitable for coherent data transmission, because they can introduce an unwanted frequency modulation (or “chirp”) that gets in the way of advanced modulation. This effect also complicates modulator biasing and operation.
Previous studies, including research from Bell Labs, have proposed mitigating the chirp through a mirrored structure in which a pair of MRMs operate in counterbalance. While
promising, this approach had, until now, never been demonstrated in a complete coherent transmitter employing advanced modulation formats at high speeds.
However, the rapid growth of AI has intensified the race to deploy MRMs in the demanding data centre environment. To develop a better understanding of MRMs, we conducted an extensive study of their dynamics [1] to determine how to fully exploit their potential.
Figure 1: Illustration of the ultra-compact silicon photonic modulator integrating microring modulators, modulating both the amplitude and phase of light. It enables fast connectivity for electronics through co-packaged optics technology
Building on this, we have now proposed an MRM-based transmitter capable of performing complex modulation. As illustrated in Figure 1, our design nests two counter-balanced MRM structures within a higher-level Mach-Zehnder interferometer configuration. Complex modulation encodes data in two dimensions—in phase and quadrature with a dedicated counter-balanced MRM structure handling each data stream.
Our design effectively eliminates the drawbacks of MRMs to allow advanced modulation at unprecedented speeds. Harnessing the energy-efficiency and compactness of this technology, our proposed device miniaturises coherent silicon modulators and eases the challenges of co-integration with advanced electronics. Compared to traditional travelling wave Mach-Zehnder modulators, our approach offers significantly higher modulation energy efficiency and a much smaller footprint.
Furthermore, thanks to the natural wavelength selectivity of MRMs, the proposed architecture is inherently compatible with WDM. This feature is particularly valuable for implementing multi-wavelength transmitters driven by optical frequency combs, eliminating the need for additional multiplexing or demultiplexing components.
In rigorous laboratory tests conducted at the Centre for Optics, Photonics and Lasers (COPL), our device achieved staggering performance metrics, with data transmission rates exceeding 1T (terabits per second) at 180 Gbaud over distances of up to 80 km. This marks the fastest MRM-based transmission reported to date, achieved with
exceptionally low modulation energy consumption – just 10.4 femtojoules per bit.
Owing to its compact on-chip footprint, we also demonstrated a record shoreline bandwidth density exceeding 5T per millimetre. This metric is particularly significant, as it underscores the device’s suitability for CPO systems, enabling tight integration with advanced electronic chips.
Achieving such record-breaking performance was not without challenges. Operating MRMs at their resonance wavelength while using high laser powers requires specific
stabilisation methods, which have been underexplored in previous research. Additionally, evaluating the novel device at these speeds demanded state-of-the-art laboratory setups. However, our interdisciplinary team and cutting-edge lab facilities enabled us to overcome these obstacles.
Figure 2: The authors are: (standing, from left to right) Simon Levasseur, Zibo Zheng, Farshid Shateri, Alireza Geravand. (seated, from left to right) Leslie A. Rusch, Wei Shi.
Vision for the future
As well as being fascinating in its own right, our new coherent silicon photonic modulator is a significant technological advancement for AI infrastructure, since it enables processors to communicate as if they were only centimetres apart, even across large datacentres spanning tens of kilometres. Dramatically reducing the size and
power consumption of coherent optics enables datacentres to expand their cluster sizes by employing coherent optical interconnects.
Furthermore, coherent optics enables dynamic network optimisation within datacentres through optical switching technology, thanks to the additional link budget provided by utilising coherent links. This flexibility allows physical networks to adapt rapidly, optimising resources based on workload demands – crucial for next-generation AI hardware.
On the manufacturing side, our silicon photonic modulator technology is fully compatible with standard semiconductor fabrication methods, enabling economical mass production. However, several hurdles remain before widespread adoption will become practical. Chief among them is reducing the complexity of the digital signal processing required for coherent detection. Promising pathways include operating in the less dispersive O-band [2] and integrating analogue coherent technologies to minimise reliance on energy-intensive digital processing.
Our lab has a strong track record of advancing silicon photonics from concept to widely adopted practice. Nearly a decade ago, we published the first demonstration of high-speed multi-level modulation in silicon MRMs – overcoming early scepticism and inspiring subsequent research efforts. This technique has since been adopted in commercial AI products, including recent implementations by companies such as NVIDIA.
Figure 3: A packaged MRM-based coherent modulator undergoes high-speed data transmission experiment at Université Laval’s COPL laboratories.
Looking ahead, our research opens pathways for new architectures of compact coherent interconnects. Future devices based on this breakthrough have the potential to
transform how data transmission systems are constructed by facilitating the integration of coherent optical technology and electronics. Scalable, energy-efficient coherent optical interconnects could usher in an era in which vast clusters of processors, memory, and storage systems are seamlessly integrated, enhancing AI performance while reducing environmental impact.
As silicon photonics continues to mature rapidly, innovations like ours will play a pivotal role in shaping the future landscape of high-performance computing and AI infrastructure.
