In recent years, as the pace of Moore’s Law has decelerated, there has been a growing interest in exploring novel hardware architectures built upon emerging device technologies. This shift is driven by the limitations of traditional silicon-based electronics in meeting the increasing demands for higher performance and energy efficiency. Among the various innovative solutions being investigated, photonic accelerators have emerged as particularly promising candidates. Photonic accelerators leverage the principles of photonics to process information at the speed of light, offering several distinct advantages over conventional electronic counterparts. One of the most significant benefits of photonic accelerators is their exceptionally high operating frequency, which enables ultra-fast data...[ Read more ]

In recent years, as the pace of Moore’s Law has decelerated, there has been a growing interest in exploring novel hardware architectures built upon emerging device technologies. This shift is driven by the limitations of traditional silicon-based electronics in meeting the increasing demands for higher performance and energy efficiency. Among the various innovative solutions being investigated, photonic accelerators have emerged as particularly promising candidates. Photonic accelerators leverage the principles of photonics to process information at the speed of light, offering several distinct advantages over conventional electronic counterparts. One of the most significant benefits of photonic accelerators is their exceptionally high operating frequency, which enables ultra-fast data processing and transmission. This characteristic is crucial for applications requiring real-time data handling and high-throughput computing. Moreover, photonic accelerators are renowned for their low power consumption. Unlike electronic devices, which generate substantial heat and require extensive cooling mechanisms, photonic devices operate with minimal energy dissipation. This efficiency not only reduces operational costs but also aligns with the growing emphasis on sustainable and environmentally friendly computing solutions.

This dissertation explores the design and implementation of high-performance domain-specific photonic accelerators, focusing on applications in deep neural networks, fully homomorphic encryption, hyperdimensional computing, and graph processing. By leveraging the high bandwidth and low latency characteristics of photonics, we have designed several innovative photonic accelerator architectures that significantly enhance computational performance and energy efficiency.

Firstly, we introduce PHANES, a ReRAM-based photonic accelerator for deep neural networks. PHANES performs multiplications within ReRAM and parallel weighted accumulations during optical transmission, thereby reducing the need for ADC/DAC converters. Additionally, we propose a progressive bit-depth technique to address the memory wall problem. Evaluations demonstrate that PHANES improves energy efficiency by 6.09x and throughput density by 14.7x compared to state-of-the-art designs.

Secondly, we present PhotonNTT, the first photonic Number Theoretic Transform (NTT) accelerator designed to address the computational challenges of fully homomorphic encryption (FHE). By formulating NTT into matrix-vector multiplication (MVM) operations and employing parallel photonic MVM units, PhotonNTT achieves a 50x improvement in throughput and a 63x enhancement in energy efficiency over existing NTT accelerators.

Thirdly, we propose OpticalHDC, an ultra-fast photonic hyperdimensional computing (HDC) accelerator. This accelerator incorporates configurable dot product cores based on microring resonators, which can be dynamically reconfigured for either encoding or classification stages. OpticalHDC achieves a significant speedup, with up to 47.87x and 177.37x improvements compared to state-of-the-art ASIC chips and GPUs, respectively.

Finally, we introduce GRAPE, a photonic graph processing accelerator featuring an optical memory interconnection network (OMIN). GRAPE utilizes photonic engines for ultra-fast MVM computations and a customized data mapping scheme for efficient graph algorithm execution. It demonstrates a 151x performance improvement and a 1111x energy efficiency enhancement over GPU platforms, and outperforms state-of-the-art processing-in-memory (PIM) accelerators by 207x in performance and 114x in energy efficiency.

Overall, this dissertation presents a comprehensive study on the development of photonic accelerators for various computationally intensive domains, showcasing their potential to overcome existing limitations in conventional electronic architectures.