Extended Data Fig. 6: Cascaded computing.
From: Synthetic-domain computing and neural networks using lithium niobate integrated nonlinear phononics
Our computing units can be cascaded by interleaving sum frequency operation mode (SFOM) and difference frequency operation mode (DFOM). This process implements the multiplication of three matrices, Y=U−1 U V. First, the multiplication W=UV is performed using SFOM, where the inputs U and V are encoded in the fundamental frequency band, and the output W shows up in the second-order frequency band. The output W is routed into DFOM as an input signal X, for the operation Y = U−1 X. Note that the cascaded signals remain at the same frequency bins and no digital signal processing is needed. By encoding U−1 into the fundamental frequency band as another input, the output Y is obtained at the fundamental frequency band. Numbers are rounded to two decimal places. The measured Y closely matches the input V, demonstrating a high fidelity in cascaded computing. We note that we detected and regenerated the analog signal in cascade, but no additional digital processing is applied; this is due to the technical limitations of our experimental setup that prevent simultaneous measurement of two devices.
