1. What are the inherent architectural deficiencies of linear optical circuitry architectures based on coherent interferometric meshes?
The inherent architectural deficiencies of linear optical circuitry architectures based on coherent interferometric meshes originate from the accumulating nature of fabrication imperfections across their cascaded nodes. These deficiencies impact size scaling and overall performance credentials, including fidelity, programming complexity, processing bandwidth, energy-efficiency, and footprint-efficiency. Specifically, these architectures suffer from insertion loss (IL) that scales nonlinearly to the loss of the constituent matrix node, high-complexity programming, limited computing nodes' update speed, reduced and nonrestorable fidelity due to differential optical path losses, and an application portfolio confined to unitary transformations. The universalization of these architectures requires the adoption of Singular Value Decomposition (SVD) schemes, which exacerbate these disadvantages. However, advancements in photonic integration technology have led to ultra-low loss 2x2 MZI technology, facilitating progress in scalability and IL performance, although not achieving high-bandwidth operation with this low-loss envelope. Despite improvements in programming complexity and mesh calibration through self-configured linear optics, fidelity restoration and accurate mapping of linear transformations remain native architectural drawbacks. The paper presented in the section introduces a novel SiPho linear operator that achieves perfect on-chip universal linear optics, breaking through the IL-fidelity-scale tradeoffs of MZI-mesh based approaches. This performance is enabled by a distributed tree-based power split-and-recombination stage and a bijective weight mapping, as demonstrated by the fabricated 4x4 SiPho Xbar architecture. The experimental record-high on-chip fidelity of 99.997+-0.002% in implementing 10000 arbitrary linear transformations highlights the fidelity restoration capabilities and scalable, low-complexity programming of this architecture.
read more
2. How does the Xbar architecture operate as a vector-matrix linear operator?
The Xbar architecture operates as a vector-matrix linear operator by directly mapping the scalars of an N-elements-long vector X and an NxM matrix W into individual photonic modulating components. This is achieved through the coherent recombination of modulated light via MZIs (Multi-Mode Interference Devices) that are nested into splitting and recombining tree configurations. The N-elements of the input vector X are encoded to equivalent modulators located at the #N outputs of a 1:N splitter, with the resulting modulated light being equally shared among the #M Xbar columns. Each modulator at each row of each column weighs the respective element of X, according to the corresponding element of W. The weighted vector's elements of the #N rows of each column are summed via an N:1 combiner, completing the required linear transformation in the electric field. This process allows the Xbar architecture to perform the necessary linear transformation efficiently and accurately.
read more
3. How does the Xbar architecture compensate for fabrication imperfections?
The Xbar architecture compensates for fabrication imperfections by utilizing a bijective weight-to-node mapping that ensures high-accuracy calibration. This mapping takes advantage of the absence of interdependence between the matrix nodes of different columns. Additionally, a low complexity O(N) hardware-aware (HA) programming algorithm is developed to program the Xbar. The algorithm uses a testbed to program the 1st column of the fabricated SiPho Xbar. By biasing the node-EAMs of the 4 rows at discrete bias levels, the column output can be accurately measured and recorded. The linear transformations are then approximated using a linear regression algorithm, which accounts for the unavoidable fabrication variation-induced discrepancy between the constituent and reference values. This approach allows for precise mapping of the transformation matrix elements to the silicon photonic Xbar, ensuring fidelity transformations despite fabrication imperfections.
read more
4. How does inter-column fidelity restoration work?
Inter-column fidelity restoration compensates for fabrication imperfections by balancing losses between columns. It involves selecting the column with the lowest output power and enforcing attenuation factors at the outputs of other columns. This process forms a loss-balanced Xbar output vector, increasing achieved fidelity from 94.35% to almost unity fidelity of 99.997%. The approach is based on the functionality of the fidelity restoration mechanism of the Xbar photonic linear processor, which aligns with expected performance variations due to fabrication variations in photonic building blocks. The schematic representation in Fig. 3 (a) shows the enforced attenuation factors and the resulting column output power distributions after fidelity restoration. This programming significantly enhances the achieved fidelity, with the statistical error induced by the measurement procedure converging to the error margin.
read more