Interconnects / analysis
Co-packaged optics brings an operating-model question to AI networking
The 2026 sampling, production and supply announcements make serviceability, energy boundaries and replacement responsibility central to evaluation.
Search Optical AI Stack
7 articles
Compute / news
FLARE brings memory into a photonic-computing research architecture
The August paper addresses stored parameters and intermediate states; the available evidence does not support a whole-system efficiency ranking.
Interconnects / news
Lightmatter's 1.6 Tbps-per-fiber sampling targets interconnect density
The March Passage announcement concerns data movement; its evaluation platform keeps measured bandwidth separate from efficiency targets.
Interconnects / news
NVIDIA puts Spectrum-X Photonics in production, with deployment evidence still separate
The May announcement advances co-packaged optical networking while leaving shipment timing and workload gains as distinct questions.
Benchmarks / guide
How to compare an optical-AI benchmark
A benchmark comparison needs compatible tasks, precision, accuracy targets, baselines, and energy boundaries.
Benchmarks / guide
Where an optical AI energy number begins and ends
An efficiency figure needs a workload, precision, and system boundary before it can describe real computing impact.
Interconnects / guide
Optical interconnect is not optical compute
Moving data with light and performing a computation with light are related architectures, not interchangeable claims.