Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 5: Poster
AKPIK 5.8: Poster
Thursday, March 12, 2026, 15:00–16:30, P5
Enabling high performance analog photonic computing using SFP transceivers — •Arvid Gansäuer1, Mingwei Yang1,2, Okan Akyüz1,2, Lennart Mannteuffel1, Felix Kübler1, Konrad Tschernig1, Enrico Stoll1, and Janik Wolters1,2 — 1Technische Universität Berlin, Berlin, Germany — 2Institute of Space Research, German Aerospace Center (DLR), Berlin, Germany
Photonic analog processors promise energy-efficient, parallel computing, specifically to tackle future machine learning and artificial intelligence (ML/AI) workloads. The fundamental mathematical operation of ML/AI computations, vector-matrix multiplication, is naturally suited to be performed in an optical setting [1]. However, many approaches utilize specialized, custom-made light sources and modulators to encode input vectors [1,2]. In this work we use commercial 1550 nm SFP transceivers to encode these vector inputs. We employ the incoherent excitation approach, which enables the use of light pulses generated by independent transceivers without any phase stabilization. To encode vector elements as incoherent light amplitudes, we generate sequences of 0- and 1-pulses within a time bin shorter than the integration time of the measuring photodiode. Using an FPGA, we achieve parallel transmission of photonic signals via 4 transceivers at 1.25 GHz/N, where N is the number of intensity levels to approximate the analog signal. Thus, our approach enables the use of robust, readily available SFP-transceiver modules for high-performance analog photonic computing. [1] Y. Shen et al. Nat. Photon. 11, no. 7, p. 441 (2017), [2] J. Feldmann et al., Nature 589, no. 7840, p. 52 (2021).
Keywords: Photonic Computing; Optical Inference; SFP-Transceivers; High Performance Analog Computing; Input Data Vector Encoding