Nature paper demos integrated photonic neural network trained end-to-end with on-chip backprop

jwt0625 · x · 2026-09-27

• Ashtiani et al. published in Nature an integrated photonic deep neural network trained end-to-end via on-chip gradient-descent backpropagation, with all linear and nonlinear computations performed on a single photonic chip.
• Previously, photonic neural network training relied on digital computers to run backprop (degrading under device-to-device and environmental variation) or gradient-free algorithms. The key breakthrough is a scalable on-chip activation gradient.
• The poster adds a hands-on layout critique: much of the chip is probe pads rather than routing, with high-speed GSG pads and tapering reused as-is, and links a review on integrated optical phased arrays (Liu 2026).

Original post →

More from Infra

• Running NixOS on DGX Spark: open-source project ships USB images and Nix module — _nateraw · 2026-09-27
• ninfer-ext: a community fork adding faster inference and Qwen3.8-Flash support — giveen · 2026-09-27
• MiniMax H3 runs locally on an RTX 3060: 20-second 540p clip took 15 minutes — Adventurous-Rich-965 · 2026-09-27
• RAM prices up 500%: buy local AI hardware now, says analyst roundup — julianharris · 2026-09-27
• UK grid power shortfall could delay a major AI project by years — rvp · 2026-09-27
• FT: OpenAI projects negative free cash flow of $278B through 2030 on compute spending — Beth_Kindig · 2026-09-27

Infra channel →