Broadband photonic tensor core with integrated ultra-low crosstalk wavelength multiplexers

Abstract The integration of artificial intelligence (AI) systems in the daily life greatly increases amount data generated and processed. In addition to large computational power required, hardware needs be compact energy efficient. One promising approach fulfill those requirements is phase-change material based photonic neuromorphic computing that enables in-memory computation a high degree parallelization. following, we present an optimized layout tensor core (PTC) which designed perform real valued matrix vector multiplications operates at telecommunication wavelengths. We deploy well-studied Ge 2 Sb Te 5 (GST) as optical attenuator single positive multiplications. order generalize multiplication arbitrary factors, develop novel symmetric unit directly includes reference-computation branch. variable GST modulation depth dB over wavelength range 100 nm with dependency below 0.8 dB. passive circuit itself ensures equal coupling main-computation branch complete range. For first time, integrate multiplexers (MUX) together crossbar array on-chip, paving way towards fully integrated systems. MUX are crucial for PTC since they enable multiple channels array. minimize crosstalk between by designing Bragg scattering MUX. By cascading, achieve extinction ratio larger than 61 while insertion loss 1