By means of artificial intelligence (AI), the amount of data processed has experienced an unprecedented surge, necessitating advanced computational techniques and technologies. Simultaneously, the hardware responsible for processing this data must demonstrate energy efficiency while maintaining a compact design. Phase change material based photonic neuromorphic computing addresses these challenges by enabling energy efficient and fast in-memory computation with a high degree of parallelization. The input data, encoded optically, is channeled into an on-chip matrix multiplication setup capable of executing parallel Multiply-Accumulate (MAC) operations using multiple wavelengths. Central to this computation are the PCM cells, which alter their refractive index according to their crystalline state[1]. These PCM cells function as non-volatile on-chip memory units. Typically, the state of these PCM units is established optically, with additional in-plane inputs required for each matrix cell. However, our approach seeks to set the state of the PCM cell using a out-of-plane vertical-cavity surface-emitting laser (VCSEL) array positioned atop the photonic-logic matrix. This approach significantly enhances the design flexibility for large scale matrix-vector multiplications and therefore opens up new possibilities for efficient computation.
|