Silicon photonic neural networks hold promise to transform artificial intelligence by enabling faster training and inference with significantly reduced power consumption. This potential leap in efficiency could revolutionize data centers, high-performance computing, and edge computing, minimizing environmental impact while expanding the boundaries of computational possibilities. At LSIP, an energy-efficient, heterogeneous III-V/silicon photonics platform has been developed as the underlying device-level foundation for innovative neuromorphic computing architectures.
Within our technology platform, quantum dot based active devices such as lasers and photodetectors are integrated on the same wafer as passive devices such as waveguides and microring resonators. With these optoelectronic devices, we are able to physically instantiate each fundamental building block of artificial neural networks all on the same chip. For instance, low-power, high-speed memristors are integrated within this platform for the storage of weights, enabling the training of deep neural networks on-chip. Furthermore, non-linear activations can be implemented using a variety of active optoelectronic devices such as quantum dot APDs and lasers, eliminating the need to shuttle data off of the chip in between neural layers. With these elements, we are building an energy-efficient neuromorphic hardware platform for deep learning and AI on the edge.
Recently, interest in programmable photonics integrated circuits has grown as a potential hardware framework for deep neural networks, quantum computing, and field programmable arrays (FPGAs). However, these circuits are constrained by the limited tuning speed and high power consumption of the phase shifters used. We have discovered memresonators, or memristors heterogeneously integrated with silicon photonic microring resonators, as phase shifters with non-volatile memory. These devices are capable of retention times of 12 hours, switching voltages lower than 5 V, an endurance of 1,000 switching cycles. Also, these memresonators have been switched using 300 ps long voltage pulses with a record low switching energy of 0.15 pJ. Furthermore, these memresonators are fabricated on a heterogeneous III-V/Si platform capable of integrating a rich family of active and passive optoelectronic devices directly on-chip to enable in-memory photonic computing and further advance the scalability of integrated photonic processors.
Heterogeneous III-V-on-silicon photonic integration has proved to be an attractive and volume manufacturable solution that marries the merits of III-V compounds and silicon technology for various photonic integrated circuit (PIC) applications. The current main-stream Ethernet trends for larger bandwidth are pushing higher modulation baudrate or employing advanced modulation format for datacom applications. However, neither solution is likely able to significantly drive overall solution cost and energy efficiency to the best sweet spot, nor to unfold the full potential of heterogeneous integration. Here we review our innovations on a special heterogeneous III-V-on-silicon platform, and the development of a dense wavelength division multiplexed (DWDM) transceiver. A 40-channel DWDM architecture and platform fabrication are discussed first, followed by experimental demonstration of each high-quality building block. InAs/GaAs quantum dot material is the choice for building robust multi-wavelength lasers, amplifiers, and high-speed avalanche photodetectors (APDs) which complements the more mature SiGe APDs. A metal-oxide-semiconductor capacitor phase shifter is a mission critical structure to provide athermal and efficient tuning for deinterleavers and microring resonators, and high-speed modulation. A successful 8 × 25 Gb/s link demonstration paves the way for the world’s first fully-integrated DWDM PIC on Si with Terabit/s aggregated bandwidth and energy efficiency likely to ∼100 fJ/bit.