Data growth and AI demand low-latency hardware. ML is 90% convolutions (conv). (N-K), where N is input and kernel matrix size (assuming squared matrices). Classification ML kernel-weight changes are infrequent. Photonic ASICs benefit from compute-in-memory. Photonic memory accesses in 0.3ns and uses 100fJ/access, 100x less than SRAM. PCM photonic memory needed photon production and detection. C-band PIC detectors require 50nW above 30GHz. Assuming 1% laser wall-plug efficiency and 2dB per coupler, PCM-written P-RAM memory READ energy is 1fJ/access for OOK signal at 30GHz data rates, or 10fJ/access for greater bit resolution. Photonic connections 100x faster than SRAM. Off-chip learning kernels can execute conv-underlying MAC operations on electro-optic PICs. Specifically, utilizing fiber optical-based discrete components (non-PIC based) and PIC-based demonstrations show the possibility for efficient photonic MAC and hence conv. acceleration.
|