What question did this study set out to answer?

This research examines photonic computing as a more energy-efficient alternative for neural network processing in AI applications.

May 17, 2026Open Access

Photonic Computing Architectures for Accelerated Artificial Intelligence: A Post-Silicon Paradigm for Energy-Efficient Neural Network Processing

Key Points

This research examines photonic computing as a more energy-efficient alternative for neural network processing in AI applications.
Evaluated three photonic computing architectures through a three-phase mixed-methods design.
Phase 1 benchmarked photonic and electronic AI accelerators for energy efficiency.
Phase 3 conducted multi-criteria decision analysis on technology readiness and investment.
Phase 1 revealed a median energy efficiency advantage of 2.36× (5.19 vs. 2.20 TOPS/W) for photonic over electronic accelerators.
PCM tensor cores ranked highest in multi-criteria decision analysis with a score of 6.72/10, surpassing MZI and MRR.
Challenges include thermal drift and accuracy degradation affecting practical application of architectures.

Abstract

The exponential growth of artificial intelligence workloads has created an unsustainable trajectory in computational energy consumption, while the deceleration of Moore’s Law and breakdown of Dennard scaling constrain electronic hardware from meeting this demand. This dissertation investigates photonic computing as a post-silicon paradigm for energy-efficient neural network processing, evaluating three architectures—Mach-Zehnder interferometer (MZI) mesh processors, micro-ring resonator (MRR) weight banks, and phase-change material (PCM) tensor cores—through a three-phase mixed-methods design. Phase 1 benchmarked four photonic against four electronic AI accelerators, revealing a median energy efficiency advantage of 2. 36× (5. 19 vs. 2. 20 TOPS/W) with a large effect size (Cohen’s d = 0. 76) and directional support from 89. 8% of 10, 000 bootstrap resamples. Statistical significance was not achieved (p =. 354) owing to 14. 9% power in the near-census sample (n = 4 per group), consistent with a Type II error. Phase 2 modeled architectural barriers: silicon MRRs exhibited thermal drift of ~100 pm/°C, MZI processors lost >60% accuracy at ±3°C deviation, 64×64 meshes required phase error tolerances of σ ≤ 0. 02 rad for 90% accuracy, and optical loss at 2. 13 dB/stage constrained practical depth to ~100–150 stages. Phase 3 conducted multi-criteria decision analysis integrating technology readiness, investment, and patent data. PCM tensor cores ranked highest (6. 72/10), exceeding MZI (5. 65) and MRR (4. 95) in a ranking robust under ±5% weight perturbation, supported by 2. 164 billion in private investment and patent growth exceeding 30% CAGR. A five-layer sanity-check audit (8 PASS, 9 CAUTION, 3 FAIL) confirmed physical plausibility and internal consistency while flagging statistical power and outlier sensitivity as limitations. The evidence suggests photonic architectures offer substantial energy efficiency potential, with PCM designs providing the most promising long-term pathway contingent on thermal management and fabrication advances, while MZI processors represent the most viable near-term bridging technology.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper