Photonic matrix core on thin-film lithium niobate runs in-situ backpropagation at 8-bit precision
jwt0625 · x · 2026-09-25
An arXiv paper by Ren et al. presents a signed incoherent optical matrix multiplier built on thin-film lithium niobate that natively supports signed inputs and weights, and demonstrates closed-loop in-situ backpropagation on a 16×16 photonic core — forward computation, nonlinear operations, error propagation, and gradient computation all run on measured physical outputs, folding real device responses and non-idealities into training.
Key results:
- 8-bit multiplication and 10-bit accumulation precision across 256 channels
- 10-bit accuracy maintained in tiled 256×256 matrix computation
- Executes Transformer linear operations in a BERT-mini workload
The authors position it as a fundamental, scalable building block for practical photonic neuromorphic computing. The accompanying thread also debates criticisms of photonic computing: slow weight loading, and the control/IO overhead of a mere 16×16 array.
More from Infra
- Dev builds local AI GTM workflow, argues the next platform entry point is hardware-bound — dotey · 2026-09-25
- US Faces Memory Chip Conundrum as AI-Critical Prices Skyrocket, WSJ Reports — pstAsiatech · 2026-09-25
- Qwen Flash Next IQ4_XS beats 27B FP8 on MMLU-Pro, GPQA and GSM8K in community eval — smallDeltaBigEffect · 2026-09-25
- High-BW TFLN Modulator Demo, but the Dual-Band Grating Coupler Steals the Show — jwt0625 · 2026-09-25
- Speculative decoding and the shift to inference as systems engineering — Arindam_1729 · 2026-09-25
- Engineer argues HBM stacks dilute bandwidth to 20% of a single DRAM layer — knowrohit07 · 2026-09-25