With the emergence of artificial intelligence, there has been a significant surge in demand for hardware performance. Stronger computing power has been achieved over the past 40 years by scaling down transistors to gain higher computing density and improving memory bandwidth to overcome communication latency. However, pushing the limits in the density and complexity of integrated circuits (ICs) has caused the traditional von Neumann computing architecture to become inadequate in supporting fields like automatic driving, big data, and the Internet of Things (IoTs). “In-memory computing” mimics the human brain’s thinking process, integrating computing functions into memory to avoid data exchange bottlenecks and transmission time decay in conventional computers. This paradigm is promising due to its low latency, parallelism, and scalability. Currently, various technological forms for in-memory computing have been proposed. Photonic in-memory computing hardware based on chalcogenide phase change materials (PCMs) combines existing dielectric materials widely used in memory technology with novel optical computing technology. Benefitting from anti-electromagnetic interference, parallelism from light’s multiphysical dimensions, zero static power consumption, high thermal crosstalk thresholds, and reduced computing time, these chips have the potential to accelerate and improve energy efficiency in data-intensive scenarios.

Our paper reviews research progress on chalcogenide PCMs, integrated devices, and optical networks for in-memory computing applications. Furthermore, we discuss challenges and future developments regarding in-memory computing devices and integrated chips. In terms of materials, chalcogenide PCMs attract attention due to their nonvolatility and optical property contrast. We analyze the principles of optimizing chalcogenide PCMs’ performance along with historical development and strategies for material improvement. Besides, we introduce the development history of application-oriented chalcogenide PCMs with low loss and large optical constant contrast, followed by several feasible strategies to further improve the properties of materials. In terms of integration processes, we classify heterogeneous integration methods between chalcogenide PCMs and waveguides as photo-induced and electric-induced schemes. Among them, the former (Fig. 2) is simpler, while the latter (Fig. 3) presents challenges due to the performance impact resulting from metal interconnection and other processes in the standard silicon photonics process. Therefore, developing wafer-level back-end-of-line (BEOL) integration processes (Fig. 4) is crucial. In terms of reconfigurable devices, we summarize recent research on photonic integrated devices based on chalcogenide PCMs for in-memory computing, including photo-induced and electric-induced devices. Devices controlled by photo-induced schemes are faster and consume less energy, while electric-induced devices suit large-scale networks with high-performance microheaters. In terms of networks and applications, we review research on in-memory computing using optical networks manipulated by chalcogenide PCMs, showing advantages in energy efficiency and integration density. Large-scale in-memory computing chips with excellent performance and robustness can be realized through collaboration with computing architecture design, advanced packaging, and photonic-electric co-integration technologies.

In summary, in-memory computing supports a series of complex, large-scale computing applications efficiently. As a novel paradigm, in-memory computing holds the promise of facilitating a multitude of applications, particularly those reliant on artificial intelligence, by offering rapid and energy-efficient solutions. This is due to optical in-memory computing architectures that utilize chalcogenide PCMs, which are specifically designed to enhance the processing of data-intensive tasks characterized by occasional configuration demands. However, challenges remain, including loss reduction in chalcogenide PCMs, improvement of endurance and multilevel precision, development of large-scale BEOL heterogeneous integration processes, and scalability of computing architectures. If the problems mentioned before are addressed, a versatile in-memory computing chip reconfigured by chalcogenide PCMs can achieve high-speed, low-power performance, fully leveraging in-memory computing advantages.