With the explosive growth of data, tensor processing has emerged as a pivotal component of the next generation of artificial intelligence (AI) algorithms. Current photonic convolutional processors transform tensor convolutions into multi-channel general matrix multiplication (GeMM), which follows the path of electronic counterparts, leading to data replication and hardware complexity. In this study, we experimentally and theoretically demonstrate a photonic tensor processing unit (PTPU) with a single modulator, which offers a more concise approach for multi-channel standard tensor convolution processing, different from the channel-wise convolution method. By executing multi-tensor parallel computing instead of multi-channel parallel computing, PTPU can directly produce feature tensors without clock synchronization and delay compensation between multiple channels and allows more release of physical hardware resources. Furthermore, an integrated array of semiconductor optical amplifiers (SOAs) are used to be photonic synapses for programmable weight bank, demonstrating record-high precision of 9.2 bits for weights. In the proof-of-concept experiment, we extracted features from a 3-channel (RGB) image in horizontal and vertical directions, using integrated multi-wavelength photonic tensor kernels. We then built a 3D convolutional neural network to predict the presence of COVID-19 based on computer tomography (CT) scan data consisting of 64-channel tensors.
Driving signal reflection on traveling wave electrodes (TWEs) is a critical issue in Mach–Zehnder modulators. Fabrication variation often causes a random variation in the electrode impedance and the signal reflection, which induces modulation characteristics variation. The variation of reflection could be intertwined with the variation of other electrode characteristics, such as microwave signal attenuation, resulting in complexity. Here, we characterize the (partial) correlation coefficients between the reflection and modulation characteristics at different bit rates. Decreasing correlation at higher bit rates is observed. Device physics analysis shows how the observed variation can be related to nanoscale variation of material properties, particularly in the embedded diode responsible for electro-optic modulation. We develop a detailed theory to analyze two variation modes of the diode (P-i-N diode or overlapping P/N regions), which reveal insight beyond simplistic diode models. Microwave signal attenuation tends to reduce the correlation with on-electrode reflection, particularly at high bit rates. The theory shows the relative importance of conductor-induced attenuation and “dielectric”-induced attenuation, with different dependence on the frequency and fabrication variation. Strategies on how to mitigate the effect of variation for better fabrication tolerance are discussed by considering three key factors: pre-shift in structural design, bias condition, and fabrication control accuracy.
20 February 2025 in Journal of Lightwave Technology
Micro-ring modulators with a small footprint and energy consumption have been explored for densely integrated high throughput optical interconnects. However, their strong dispersive response results in frequency chirp and pulse broadening, which limits their signal transmission distance. Here we explore the low chirp operation of a dual-racetrack modulator symmetrically coupled to a bus waveguide. With opposite index changes in the two resonators, the coherent couplings of two resonators' optical fields with the bus waveguide and their interference can cancel out the output phase variation due to a single resonator. This allows the push-pull operation to achieve a lower chirp compared to the single-drive operation. The advantage of lower chirp operation is experimentally verified through open eye diagram with 100 Gbps PAM-4 transmitted over 2 km single-mode fiber (SMF), with a peak-to-peak driving voltage of 1.6 V.
Current photonic convolutional processors transform tensor convolutions into multi-channel general matrix multiplication (GeMM), leading to data replication and hardware complexity. In this study, we propose and experimentally demonstrate a photonic tensor processing unit (PTPU) with single dataflow, which offers a more concise approach to multi-channel standard tensor convolution processing. In experiment, we extracted features from a 3-channel (RGB) image in horizontal and vertical directions using an integrated multi-wavelength source. We then built a $3\mathrm{D}$ -convolutional neural network to predict the presence of COVID-19 based on computer tomography (CT) scan data consisting of 64-channel tensors.