A fully-programmable integrated photonic processor for domain-specific and general-purpose computing
First single programmable photonic processor handling both NP-complete problem solving (100% accuracy) and general-purpose matrix computation (97% MNIST accuracy) without hardware modification
A fully-programmable integrated photonic processor for domain-specific and general-purpose computing
Abstract
The authors present an integrated photonic processor capable of addressing both specialized computational problems and general matrix operations. The system solves NP-complete problems (subset sum and exact cover) with complete accuracy across 2^N+ instances, demonstrates "high-precision optical dot product," and achieves 97% accuracy on MNIST image classification — all on the same physical chip without hardware modification.
Key Contributions
- Unified architecture: First demonstration of a single programmable photonic processor handling both domain-specific NP-complete problem solving and general-purpose matrix computation without hardware modifications
- NP-Complete problem solving: Successfully solves "more than 2^N different subset sum problem instances" and exact cover problems with 100% accuracy
- General-purpose computing: Implements optical dot products with 7.22-bit precision accuracy and executes multi-kernel convolution operations
- Integrated system: Custom 512-channel optoelectronic computing board providing full programmable control
- Silicon nanophotonic chip: 6mm × 5mm chip with 498 optical components using air-trench design to reduce thermal crosstalk
Methodology
Hardware architecture:
- Silicon nanophotonic chip with thermally-modulated Mach-Zehnder interferometers (MZIs), directional couplers, and grating couplers
- Air-trench design reduces thermal crosstalk between components
- 512-channel optoelectronic computing board for full programmable control
Programming approach:
- Reconfigurable MZI states (bar, cross, balanced configurations)
- Multiple input channels enable configuration of 2^N problem instances
- Abstract network representation maps computational problems to optical paths
Problem mapping:
- Subset sum: elements encoded as vertical distances; diagonal light propagation indicates element inclusion
- Exact cover: binary encoding scheme; bitwise sum detection identifies solutions
- Matrix operations: cascaded MZI networks with row-wise weight encoding
Results
| Task | Performance |
|---|---|
| Subset sum problems | 100% accuracy across 89 configurations |
| Exact cover problems | 100% accuracy |
| Optical dot product | Std deviation 0.0067 (7.22-bit precision) |
| Image edge detection | RMSE: 0.0087 |
| MNIST classification | 97% accuracy (vs. 97.5% numerical baseline) |
Limitations
- Scalability constraints: Current implementation limited to 8-dimensional dot products and 5-element target sets
- Physical constraints: ~1.5 dB/cm waveguide propagation loss; coupling loss of 4 dB per facet
- Precision limits: 7.22-bit computing accuracy may be insufficient for demanding applications
- Thermal management: Requires dedicated temperature control; crosstalk mitigation through design modifications
- Problem size restrictions: Subset sum instances limited to N=5 in demonstrations
- Signal-to-noise ratio: Requires threshold-based classification of output signals to distinguish valid from invalid results
Source: A fully-programmable integrated photonic processor for domain-specific and general-purpose computing by Feng-Kai Han et al., Shanghai Jiao Tong University; Hefei National Laboratory; TuringQ Co., Ltd.