The data shuttling between computing and memory dominates the power consumption and time delay in electronic computing systems due to the bottleneck of the von Neumann architecture. To increase computational efficiency and reduce power consumption, photonic in-memory computing architecture based on phase change material (PCM) is attracting increasing attention. However, the extinction ratio and insertion loss of the PCM-based photonic computing unit are imperative to be improved before its application in a large-scale optical computing network. Here, we propose a 1 × 2 racetrack resonator based on Ge2Sb2Se4Te1 (GSST)-slot for in-memory computing. It demonstrates high extinction ratios of 30.22 dB and 29.64 dB at the through port and drop port, respectively. The insertion loss is as low as around 0.16 dB at the drop port in the amorphous state and about 0.93 dB at the through port in the crystalline state. A high extinction ratio means a wider range of transmittance variation, resulting in more multilevel levels. During the transition between crystalline and amorphous states, the tuning range of the resonant wavelength is as high as 7.13 nm, which plays an important role in the realization of reconfigurable photonic integrated circuits. The proposed phase-change cell demonstrates scalar multiplication operations with high accuracy and energy efficiency due to a higher extinction ratio and lower insertion loss compared with other traditional optical computing devices. The recognition accuracy on the MNIST dataset is as high as 94.6% in the photonic neuromorphic network. The computational energy efficiency can reach 28 TOPS/W, and the computational density of 600 TOPS/mm2. The superior performance is ascribed to the enhanced interaction between light and matter by filling the slot with GSST. Such a device enables an effective approach to power-efficient in-memory computing.
Keywords: computational efficiency; phase change material; photonic in-memory computing; power consumption; von Neumann architecture.