2DNeuralVision now runs optical AI from a standard 19‑inch rack that plugs straight into PyTorch. The system couples a silicon‑photonics optical neural network (ONN) to a Xilinx RFSoC controller, so matrix math happens in light while developers keep using the tools they know.

This achievement demonstrates that optical computing can operate as a true drop-in AI backend, enabling developers to accelerate models using light-based matrix operations without changing their code.

Why this changes the game

• Zero barrier for ML teams: the ONN appears as drop‑in nn.Linear and nn.Conv2d layers, so existing models can offload heavy linear ops to optics with minimal or no code changes.
• Real accuracy–latency control: single‑shot runs show ~20% normalized MVM error; averaging x4 cuts that to ~11% and heavy averaging approaches ~3%, giving users a software-level dial between maximum speed and precision.
• Proof on real workloads: in high-precision operating mode, the system reaches ~98% test accuracy on MNIST (near the digital baseline) and ~72% on a CIFAR‑10 subset with the photonic core in the loop, demonstrating end-to-end interface on standard vision benchmarks.

What it means for adoption

Optical compute is a mature, deployable ML backend that Teams can prototype hardware‑in‑the‑loop workflows today, scaling seamlessly as models and datasets grow – all while dialing the precision/speed trade‑off from software. In practice, this means Europe now has a developer-ready photonic AI platform, directly supporting future demonstrators in mobility and industrial sensing with high-throughput, energy-efficient linear operations in light.