Mar 154 min read

Electronic-Photonic Systems-on-Chip: Enabling the Next Wave of Computing, Communications and Sensing

Introduction

As computing demand continues to grow exponentially, driven by applications like artificial intelligence (AI) and 5G/6G wireless communications, existing electrical interconnect technologies are hitting fundamental limits. To overcome these bottlenecks, the semiconductor industry is turning to integrated photonics - the introduction of optical components directly into advanced CMOS processes. Combining optics and electronics onto a single silicon chip unlocks new capabilities and architectures that could enable the next wave of breakthroughs across computing, communications and sensing applications [1].

CMOS-Integrated Photonics

At the heart of these electronic-photonic systems is the ability to fabricate high-performance optical components like micro-ring resonators directly in a conventional CMOS process. Figure 1 shows an example of interleaved planar PN junctions forming resonant photonic micro-rings in a standard CMOS process flow. These highly compact (diameter ~10μm), high-Q (up to 200,000) structures can be used for a variety of functions like modulation, filtering, and sensing.

A key enabler has been innovations in advanced CMOS process technology itself. Figure 2 illustrates the evolution of a 45nm CMOS process to incorporate a variety of photonic features like germanium photodetectors, silicon partial etches, nitride layers for waveguides, and v-groove fiber couplers - resulting in a monolithic electronic-photonic platform operating at the same performance as advanced pure-play CMOS transistors.

Figure 1. Resonant photonic micro-rings in silicon CMOS

Figure 2. From zero-change to standard process in 45nm

Applications and Impact

Digital Optical Interconnects for AI/ML Scale-out One of the most impactful applications of integrated photonics is in addressing the interconnect bottleneck for scaling out AI/ML workloads across large multi-chip/multi-package systems. As these models grow to trillions of parameters, the interconnect efficiency (ratio of compute time to total time including communications) drops from over 80% to below 60% when scaling out to multiple packages.

Figure 3. Interconnect limits AI scale-out runtime

Integrated photonic chiplets like the TeraPHY from Ayar Labs (Figure 4) provide a revolutionary solution - a multi-wavelength optical I/O interface integrated directly into the package alongside the compute/memory chips. Each TeraPHY chiplet provides over 4Tbps of optical I/O bandwidth across 8 ports and up to 16 wavelengths per port (Figure 5). This dramatically improves system-level throughput and efficiency.

Figure 5. Micro-ring WDM bandwidth scaling (Tx+Rx)

The TeraPHY chiplets leverage a disruptive new packaging approach (Figure 6) where high-bandwidth optical I/Os can be routed off-package through fibers and optical connectors. A SuperNova multi-wavelength laser source co-packaged with the compute chips provides the light source for communications within the rack, across racks, or to remote systems - enabling new distributed computing architectures.

Figure 6. Optical I/O enables complete physical disaggregation of distributed computing

The technology roadmap (Figure 7) shows a clear path to reaching over 250Tbps of off-package optical I/O bandwidth within the next decade by leveraging advanced packaging technologies and scaled CMOS processes. This supports future AI fabric architectures like Nvidia's proposed system with 1 million nodes that are 100x faster than today's A100 systems (Figure 8).

Figure 7. Future systems-in-package with optical I/O

Figure 8. Future AI fabric scaling: 1000x more nodes, 100x faster nodes

A key innovation enabling this scale is the SuperFabric optical interconnect architecture (Figures 9, 10) based on laser-forwarded coherent optical links. These links employ a novel coherent detection approach along with high-radix fast optical switches to provide over 100Tbps of connectivity to each package with fast reconfigurability.

Figure 9. SuperFabric for ML/AI workload scale-out

Figure 10. SuperFabric components and features

Analog/RF Optical Interconnects

Beyond digital computing, the convergence of photonics and electronics also enables transformative architectural approaches for applications like 6G/7G wireless and automotive radar sensing. The extremely high bandwidths and energy efficiencies of microring-based optical interconnects can be leveraged to disaggregate the analog/RF front-ends from the digital back-ends in systems like massive MIMO radios (Figures 11, 12).

Figure 11. Extreme massive-MIMO 6G+ radio

Figure 12. Low-power, disaggregated Rus (DRUs)

This architectural approach, combining digital optical interconnects to baseband processors with analog optical links to the distributed antenna/radar front-ends (Figure 13), dramatically reduces system power and footprint. Figures 14 and 15 show examples of implementing such analog optical links using high-performance microring modulators driven directly by millimeter-wave signals.

Figure 16 highlights the noise and linearity characteristics of single microring modulators as a function of laser power and wavelength detuning - identifying an optimal operating regime that can meet the stringent requirements of cellular and automotive radar applications. State-of-the-art prototypes have already demonstrated spurious-free dynamic range over 95dB.Hz2/3 (Figure 17), exceeding the 3GPP LTE requirement.

Figure 14. Ring-based analog millimeter-wave photonic-electronic sensing arrays

Figure 16. Single ring modulator linearity modeling

Cryogenic Interconnects and Sensing

Two other pioneering applications leverage the unique properties of integrated photonics in advanced CMOS processes. The first is an optical interconnect solution for cryogenic computing systems like quantum computers (Figure 18). Conventional electrical cabling quickly becomes impractical at scale due to thermal loads and wiring complexity.

Figure 18. Cryogenic computing (quantum and classical)

The Coherent Ultrafast Reflective Link (CURL) developed by researchers at UC Berkeley addresses this by using advanced CMOS electronics operating at cryogenic temperatures (Figures 19, 20) to drive ultra-compact microring modulators that reflect a multi-wavelength tunable laser back to a coherent receiver at room temperature (Figures 21, 22). First prototypes already demonstrate multi-Gbps link operation between 4K and 300K (Figures 23, 24).

Figure 20. 45nm SOI CMOS Cryo performance

Figure 21. Prototype Cryo-TX for 10Gbps link – subthreshold pre-amplifier

Figure 22. Prototype Cryo-TX for 10Gbps link - Micro-ring resonator modulator

Figure 23. First full CURL link demonstration

The second breakthrough application is in label-free photonic biosensing leveraging the high sensitivity of microring resonators to detect molecular and viral analytes (Figure 25). By monolithically integrating dense arrays of microrings with microfluidics and electronics (Figure 26), it enables highly multiplexed point-of-care sensing that could be extended into wearables and implantable devices for personalized healthcare applications.

Figure 25. Photonic label-free molecular bio-sensing

Figure 26. Photonic label-free bio-sensing

Conclusions

Integrated electronic-photonic systems are poised to revolutionize computing, communications and sensing by overcoming key bottlenecks with innovative new photonic device structures, link architectures, packaging and system-level solutions. While still in its early stages of commercial deployment, the field is progressing rapidly with clear roadmaps for scaling bandwidth density, energy efficiency and application scope over the coming decade.

Reference

[1] V. Stojanović, "Electronic-Photonic Systems-on-Chip for Compute, Communications and Sensing," in IEEE International Solid-State Circuits Conference (ISSCC), 2024.