Oriole Networks and AMD have deployed a photonic AI network for the UK’s Advanced Research & Innovation Agency (ARIA) Scaling Inference Lab, combining Oriole’s PRISM photonic networking with AMD Instinct GPUs and AMD EPYC CPUs. The collaboration, underway for more than a year, is aimed at testing how next-generation network fabrics can address performance, latency, and energy constraints in AI infrastructure.
Oriole’s contribution is PRISM, a photonic networking solution that replaces electronic switches in the network core with nanosecond-scale optical circuit switching. AMD is providing CPU and GPU hardware, plus technical collaboration to develop and run large-scale network models intended to be relevant to frontier-scale AI systems.
Oriole describes the ARIA cluster work as a move toward a “pure photonic AI network at scale,” with the goal of minimizing latency at the system level. For data center architects, the practical question is whether optical circuit switching can remove enough serialization, buffering, and switch power overhead in the core to materially change cluster-level efficiency—and do it in a way that’s operable at scale.
Oriole claims PRISM “cuts core power consumption by 81%” by removing electronic switches from the core, and that with photons traveling directly from chip to chip, GPU idle time can drop “from 60% today to less than 1%.” The company also says fewer electronics in the network reduces cooling demand and water usage, and that the result is higher inference throughput (“more tokens per second”) and more concurrent users served on the same hardware.
Oriole also characterizes the ARIA Scaling Inference Lab deployment as the first commercial deployment of its technology, stating it has moved “from R&D to production in just three years.” The company says its xPU-agnostic designs are now “locked” and that it’s targeting a wider rollout across the industry in 2027.
“Oriole’s AI backend networking with nanosecond optical circuit switching represents a fundamentally different way to connect accelerators at scale,” said Madhu Rangarajan, corporate vice president, Compute and Enterprise AI business, AMD. “We are helping to validate how photonic fabrics can work alongside AMD compute to deliver the low-latency, high-bandwidth connectivity that AI Inference workloads demand.”
ARIA’s Scaling Inference Lab is a testbed backed by £50m ($68m), intended to address a bottleneck in AI workloads.
Source: Oriole Networks
