.png){kind=icon}
.png){kind=icon}
Return
Reconfigurable photonic complex-valued matrix–vector multiplication processor based on time-division multiplexing
.png){kind=icon}
PRE
AI
DOI：10.1016/j.optlastec.2026.114687.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Matrix-vector multiplication (MVM) operations play a key role in photonic computing systems and signal processing. Based on a time-division multiplexing architecture, we propose a reconfigurable photonic complex-valued MVM (RP-CMVM) processor. This scheme employs complex-valued encoding and decoding techniques to perform arbitrary complex-valued operations using only three intensity modulators. Meanwhile, this architecture enables flexible switching between complex-valued and real-valued computation tasks without changing the system hardware. For computations involving any two sets of complex-valued data, the root mean square error (RMSE) of RP-CMVM output is on the order of 3E-3. Additionally, it supports parallel edge extraction in images using the complex-valued Sobel operator, with an output image RMSE of approximately 1E-2. Then, for an iris classification task, the neural network based on the RP-CMVM attained a test accuracy of 96.67%. In addition, we implemented the frequency offset compensation task in a coherent optical communication system based on the proposed scheme, and its bit error rate is basically consistent with that of the classical algorithm. Further, this scheme combines wavelength division multiplexing to realize parallel computing, and its computing speed can reach tera operations per second (TOPS) level. Therefore, the RP-CMVM processor offers an alternative solution for implementing a reconfigurable and efficient MVM computing platform.
Journal
O
IF:
0
Papers:
880
Citations:
0