Shanghai Jiao Tong University unveils scalable photonic convolution chip

Researchers at Shanghai Jiao Tong University have built SPIN, a silicon photonic convolution architecture that shifts scaling from spatial duplication to wavelength-domain interleaving. The team says the design cuts waveguide-length growth, supports reconfigurable kernels and task parallelism, and delivered strong MNIST test results on a 220 nm silicon-on-insulator chip.

Why it matters: - SPIN targets one of the biggest bottlenecks in photonic AI hardware: scaling convolution without blowing up chip size, delay-line complexity, and control overhead. - The architecture is designed to make compact integrated photonics more practical for high-throughput machine vision and other convolution-heavy workloads. - The work points to a path for combining optical parallelism with reconfigurability, which is harder to achieve in many existing photonic accelerators.

What happened: - Researchers at Shanghai Jiao Tong University developed SPIN, a spatiotemporal photonic interleaving network for scalable photonic convolution. - The study was published in Opto-Electronic Science on July 23, 2026. - The team used a recursive tree of cascaded optical interleavers to share delay lines across wavelength channels. - The architecture broadcasts a serialized input waveform onto multiple wavelength carriers, routes each channel through shared delay segments, applies kernel weights optically, and recovers signed results electronically after baseline subtraction. - The proof-of-concept chip was fabricated on a commercial 220 nm silicon-on-insulator platform. - The prototype used a three-stage cascaded interleaver network for eight wavelength channels. - At 49 Gbaud, the chip executed representative 2 × 2 convolution operations on MNIST handwritten digits.

The details: - SPIN reduces waveguide-length scaling from O(K²) to O(K log₂ K) by placing longer delay segments upstream where downstream wavelength channels can reuse them. - The number of actively controlled weighting elements scales as O(K). - In experiments, the optical waveforms matched digital ground truth with correlation coefficients above 0.98. - The reconstructed feature maps clearly highlighted digit contours. - Sixteen optical carriers were divided into wavelength groups so multiple image batches and convolution tasks could share the same physical SPIN core. - The paper also demonstrated a 2 × 4 convolution kernel on natural images from the USC-SIPI database. - The architecture is built to trade wavelength resources among kernel size, kernel geometry, patch parallelism, and task parallelism. - The authors estimate that a fully spectrum-enabled SPIN core could reach 29.7 TOPS. - The work received partial support from the National Science Foundation of China under No. 62341508 and the Shanghai Municipal of Science and Technology Project under No. 24JD1401500.

Between the lines: - SPIN is meant to solve a practical tradeoff in photonic computing: dense hardware is useful, but density often comes at the cost of flexibility or calibration burden. - The wavelength-domain approach suggests a different scaling strategy than simply adding more spatial paths, which could matter for future on-chip optical accelerators. - The projected performance is promising, but the system still depends on co-design of modulators, photodetectors, frequency-comb sources, calibration, and electronic interfaces. - The main scientific value is not just the convolution demo; it is the scaling method that lets one photonic core support more work without proportional growth in footprint.

What's next: - Further development will likely focus on integrating the supporting hardware needed for practical deployment, including sources, detectors, modulators, and calibration loops. - The architecture could be extended to larger kernels and broader wavelength-multiplexed workloads if the full optical spectrum is available. - The next test for SPIN is whether the scaling advantage holds in a more complete system, not just in proof-of-concept experiments.

The bottom line: - Shanghai Jiao Tong University’s SPIN offers a compact route to scalable photonic convolution by shifting complexity from space to wavelength.