Optical Computing
Silicon photonics, co-packaged optics, photonic AI accelerators, optical neural networks
Optical Computing
The optical computing field in 2025-2026 operates on two timescales. The near-term story is co-packaged optics, and it now runs on two integration topologies. The chiplet model is locked in: Ayar Labs' $500M Nvidia-backed raise, TeraPHY at 200+ Tbps/package (5x Rubin GPU bandwidth), TSMC COUPE making optical I/O accessible to any TSMC customer via UCIe, and the OIF 3.2T multi-vendor standard. Alongside it, Lightmatter's Passage M-series puts optics into an active photonic interposer (the M1000: 114 Tbps total, 256 fibers, on-package optical circuit switching), and commercialization is shifting to ASIC/packaging partnerships — Lightmatter–GUC (Jan 2026) productizes Passage 3D CPO for hyperscalers. Deployment timeline: early adopters 2026, broader adoption 2027, standard AI networking by 2028.
The longer-term story is photonic compute itself — and 2026 delivered both a reality check and a breakthrough. The reality check: UT Austin's SimPhony framework shows DAC/ADC peripheral overheads dominate photonic AI energy budgets, MZI meshes fail on Transformer workloads, and practical precision is capped near 8 bits, with the time-multiplexed crossbar the competitive architecture. The breakthrough: Nokia Bell Labs demonstrated the first on-chip gradient-descent backpropagation — all linear and nonlinear computation on one photonic chip — retiring the "photonic chips are inference-only" assumption. Hardware demonstrations validate both accuracy and density (Sydney's peer-reviewed inverse-design accelerator: ~400M params/mm², 90-99% biomedical imaging; SJTU's 498-component chip: 97% MNIST + 100% NP-complete). The live design tension is now ultra-dense fixed-function accelerators (Sydney) versus reconfigurable on-chip-trainable processors (Bell Labs).
Frontier — What's Moving Now
- CPO deployment wave underway, now two topologies — chiplet-CPO (Ayar Labs / TSMC COUPE / OIF) plus the active photonic interposer (Lightmatter M1000); commercialization via Lightmatter–GUC
- On-chip training crosses the threshold — Nokia Bell Labs' on-chip gradient-descent backprop (Nature 651) ends the "inference-only" era at lab scale
- DAC/ADC is the real bottleneck for photonic compute — SimPhony (April 2026) overturns prior component-level benchmarks
- Density milestone — Sydney's peer-reviewed inverse-design accelerator hits ~400M trainable params/mm² (fixed-function)
- Fixed-function vs. reconfigurable-trainable — the emerging design axis (Sydney ↔ Bell Labs)
- Quantum photonics diverging — SNSPDs >90% efficient but cryogenic; silicon photonics the preferred QKD platform
Concept Map
Concepts
| Concept | Sources | Evidence | Frontier | Last Updated |
|---|---|---|---|---|
| Photonic Neural Networks | 9 (papers + reviews) | Strong | Active | 2026-06-24 |
| Photonic Tensor Cores | 1 (paper) | Moderate | Active | 2026-04-05 |
| Photonic Interconnects | 5 (tech reports + analysis) | Strong | Active | 2026-06-24 |
| Co-Packaged Optics | 9 (tech reports + analysis + papers) | Strong | Active | 2026-06-24 |
| Photonic Accelerators | 6 (papers + tech reports) | Strong | Active | 2026-06-24 |
| Photonic Computing Limitations | 3 (papers) | Strong | Active | 2026-04-14 |
| Quantum Photonics | 1 (review paper) | Moderate | Active | 2026-04-14 |
Entities
| Entity | Type | Sources | Key Connection |
|---|---|---|---|
| Lightmatter | Company | 3 | Passage L200 (1.6 Tbps/fiber) + M1000 active interposer (114 Tbps) + GUC commercialization |
| Ayar Labs | Company | 3 | First UCIe optical chiplet, $500M raise, TeraPHY 200 Tbps/package |
| TSMC (Photonics) | Foundry | 3 | COUPE platform: 100 Tb/s per accelerator |
| Nokia Bell Labs | Research | 1 | First on-chip gradient-descent backprop (Nature 651) |
| GUC (Global Unichip Corp) | Company | 1 | Lightmatter's commercial Passage 3D CPO ASIC/packaging partner |
| OIF | Standards Body | 1 | 3.2T CPO standard: 51.2 Tb/s switch bandwidth |
| Q.ANT | Company | 1 | NPU 2: 30x energy reduction (vendor claim), shipping 2026 |
| University of Sydney | Research | 2 | Inverse-design nanophotonic NN: ~400M params/mm², 90-99% medical accuracy |
| UT Austin / ASU (SimPhony) | Research | 1 | SimPhony: system-level benchmarking, DAC/ADC bottleneck |
Timeline
See timeline.md for chronological developments (2019 through 2026).
Research Frontier
See frontier.md for active research directions, breakthroughs, and knowledge gaps.