What it is
A D2NN is a stack of diffractive optical elements (DOEs) where each layer applies a programmable phase transformation to an incoming light wavefront. The composition of multiple layers computes complex optical transformations, originally for classification (image → class) and more recently for arbitrary information routing.
Trade-off: SLM-class switching speed (single-µs to ms depending on technology — liquid-crystal SLMs slowest, LCOS/DMD faster) bounds the architecture’s applicability. Sub-ns electrical switching is not in reach.
Lineage
The lineage is genuinely a 6-year research arc with peer-reviewed milestones. The pedigree is real. The commercial productisation question is separate.
What it could enable in AI infrastructure
The Zigzag application repurposes D2NN as an optical interconnect fabric for MoE all-to-all collectives:
-
Native optical multicast without splitter losses. Conventional optical multicast uses a 1×N power splitter (12dB loss minimum for 1×16). D2NN does multicast via constructive interference at multiple output ports directly. No splitter, no fixed pattern, programmable.
-
Combined wavelength + spatial routing in one device. Conventional OCS does spatial routing only; wavelength routing requires AWG/WSS in series. D2NN handles both in one stack.
-
Per-layer fabric reconfig if SLMs hit single-µs. Today’s mechanical OCS (Google Apollo) sets topology once per workload. D2NN could potentially reconfigure mid-batch to match current expert-routing patterns.
-
No foundry queue. Free-space optical elements don’t need a wafer process. Both a moat (no foundry capacity constraint) and a manufacturability risk (alignment + thermal stability + scaling).
Competitive landscape
D2NN-based OCS is one of four credible architectures competing for the MoE all-to-all interconnect prize:
Reading list
- Lin & Ozcan 2018 (Science) — original D2NN architecture paper
- Yildirim et al. 2024 (Nat. Photonics) — nonlinear D2NN
- Dinç et al. 2024 (arXiv) — D2NN as multicast OCS