Research & Papers
In preparation
SPEM: A Split-Timescale Coupled Error Model for Stochastic-Photonic GEMM
Predicting irreducible noise floors and bitstream saturation in stochastic-photonic GEMM.
- When
- Feb 2026 – Expected Dec 2026
- Role
- First author
- Venue
- Undergraduate Research Program, Tsinghua University
- Authors
- Peilin Li
- Headline result
- 1.96% median prediction error
- Developed SPEM, a split-timescale model coupling stochastic-sampling variance with microring drift, shot noise, and WDM crosstalk to predict irreducible noise floors and bitstream saturation in stochastic-photonic GEMM.
- Built a disjoint calibration and Monte Carlo validation pipeline mapping measurable hardware parameters to GEMM-level NFMSE.
- Achieved 1.96% median prediction error over 200 held-out configurations and exact optima on a 234-point design grid; reduced sampling energy by up to 97.8% under matched error budgets and derived thermal-drift guard bands for a first-author journal manuscript.
Tech stack
- Photonic Computing
- Monte Carlo
- Error Modeling