LightIn: Versatile Silicon Integrated Photonic Processor for AI Clusters

Abstract

The LightIn versatile silicon integrated photonic processor (arXiv 2504.01463, April 2026) integrates 160 photonic components to realize multiple functions in a single platform: computing acceleration, signal processing, network switching, and secure encryption — leveraging silicon photonics for high speed, low latency, and large bandwidth. It is the first to demonstrate this breadth of AI-cluster-relevant functions on one CMOS-compatible chip.

Key Contributions

• 160-component integration in a single silicon photonic processor — large by photonic-integration standards.
• Multi-function platform: compute + signal processing + switching + encryption in one substrate.
• AI cluster relevance: positions photonics as a horizontal substrate for the data center — not just an accelerator.
• CMOS-compatible: integration with mainstream silicon process flow lowers manufacturing barriers.

Methodology

The paper reports design + fabrication + characterization of a 160-component silicon photonic chip, with demonstrations of each function (matrix-vector multiplication, optical filtering, switching, key-distribution-based encryption) on the same hardware via reconfiguration.

Results

• Functional demonstrations across all four target functions on the same chip.
• Performance claims competitive with digital-electronic counterparts on per-task metrics (specifics in full paper).
• 160-component scale advances state-of-the-art for general-purpose silicon photonic processors.

Limitations

• Per-function performance vs specialized digital ASICs (matrix multiplication on H100, switching on dedicated optical switches) requires direct benchmarking.
• 160-component scale is research-validated; production yields at this complexity not addressed.
• Full system integration with electronic AI clusters (host CPU, GPU, networking) requires further work.

Full Content

LightIn's positioning is platform-level rather than accelerator-level. Most photonic computing efforts target a specific function (matrix multiplication for AI inference, optical interconnect for chip-to-chip, all-optical switching for networking). LightIn argues for a horizontal substrate that handles multiple functions on the same chip via reconfiguration.

For AI clusters, this vision implies a different architectural pattern: photonic processors as flexible accelerators alongside GPUs, handling specific phases of the inference pipeline (pre-processing, signal handling) where photonic latency/bandwidth advantages matter most. CMOS-compatibility is the practical lever — without it, integration with the digital AI stack is prohibitive.

This complements the 2026 photonic computing wave: imec rack-unit production (PyTorch-integrated), Ashtiani on-chip backprop (Nokia Bell Labs), SKYLIGHT 3D WDM tensor core, Lightmatter production transformers, Q.ANT NPU 2. Each addresses a different gap: production-readiness, on-chip training, scaling, model fidelity, multi-function. LightIn fills the multi-function gap.

Source: arXiv 2504.01463 — Versatile silicon integrated photonic processor: a reconfigurable solution for next-generation AI clusters, April 2026