The impulse response becomes the kernel
A peer-reviewed experiment at the Chinese University of Hong Kong uses one microring resonator, an MRR, as a temporal convolution kernel. The kernel is the ring's impulse response. A single MRR at 128 Gbaud reaches 5.12 trillion operations per second. A resonance linewidth of about 10 GHz holds scalar weighting, and the result is 160 times that baseline.[1]
Scalar weighting stays at the linewidth because the ring is asked for one number. The impulse response keeps a short fading trace, and convolution uses that trace. An alternative reading counts the 160-fold figure as arithmetic inside the kernel, while task accuracy sits in the later 3 tests.[1]
The multiplexed engine stays on the bench
Wavelength multiplexing and space multiplexing build a multi-channel MRR convolution engine at 120.8 trillion operations per second. The density is 48.05 trillion operations per second per square millimetre. The same experiment checks 3 tasks: optical modulation format identification, network anomaly detection and image classification.[1]
The delivered test is these 3 laboratory tasks. An independent chip, a published accuracy beside the operation count, and a power figure sit outside this result. Delivery of the bench engine is that trio of tests.[1]