Photonic AI chip can help calculate its own corrections via in-situ gradient descent
bravo_abad · x · 2026-10-06
Zhou and colleagues demonstrated a self-correcting photonic AI chip. Chips that perform well in simulation can give wrong answers once fabricated, as tiny manufacturing errors alter how light travels through them.
Method
- Light carrying the input is sent forward through the chip, then a signal representing the output error is sent backward
- The amplitude and phase of both signals are measured inside the device
- These measurements enable in-situ gradient descent, adjusting chip settings using its actual physical response — the imperfections causing the errors are present in the signals used to correct them
Results: Across four optical wavelengths, training reduced the mean-squared error of the chip's matrix calculations from 0.53 to about 0.0013, showing that access to gradients for physical systems can be a matter of exploiting the system's own physics.
More from Infra
- Disco a dangerous short after HBM despec news; H2 China demand may flip — zephyr_z9 · 2026-10-06
- TensorFold bonds dual Thunderbolt 5 links for 83% throughput boost on Apple Silicon — AIFlow_ML · 2026-10-06
- Dual RX 6800 running Qwen3 27B: Linux + ROCm beats Windows Vulkan, 45 tok/s TG — SaGa31500 · 2026-10-06
- DeepSeek open-sources DeepGEMM, a clean and efficient GPU BLAS kernel library, now 8.5k stars — deepseek-ai · 2026-10-06
- AMD's Lisa Su predicts "very high" chip demand for years despite AI safety concerns — AryHHAry · 2026-10-06
- Budget Hybrid Local Box for Qwen3.8-Flash-Next 4-bit: 96GB RAM Plus Used RTX 3090 — Confident-Truth3607 · 2026-10-06