IN Brief:
- Oriole Networks is integrating PRISM photonic networking with AMD Instinct GPUs and EPYC CPUs.
- The deployment will run inside the £50m ARIA Scaling Inference Lab.
- Photonic networking is moving from research infrastructure toward commercial AI system validation.
Oriole Networks is deploying its PRISM photonic networking platform with AMD Instinct GPUs and EPYC CPUs inside the UK Advanced Research and Invention Agency’s Scaling Inference Lab.
The collaboration places the UK startup’s optical network fabric into a £50m facility designed to test technologies that can improve large-scale AI inference. Oriole and AMD have worked together for more than a year, with the installation intended to validate how photonic switching can reduce latency, increase throughput, and improve accelerator utilisation at cluster scale.
PRISM uses nanosecond optical circuit switching to route data directly as photons, replacing parts of the electronic switching layer normally used in data-centre networks. Oriole says the approach can reduce network-core power consumption by 81% and cut GPU idle time from around 60% to below 1%.
The deployment is Oriole’s first commercial installation after three years of development. Its xPU-agnostic architecture has now been finalised, with wider industry deployment planned from 2027 across accelerator platforms from multiple vendors.
AI infrastructure has pushed processor, accelerator, and memory performance sharply upwards, yet many system limits now sit outside the main compute package. Memory bandwidth, rack power, cooling, optical links, switching latency, power conversion, and workload scheduling all affect how much useful inference can be delivered from installed hardware.
The power layer is already changing in response, with 800VDC auxiliary power designs being developed for AI data-centre architectures. Oriole’s deployment addresses the network side of the same problem: once compute hardware becomes expensive, dense, and power hungry, poor interconnect efficiency becomes a direct cost rather than a background infrastructure issue.
Photonic networking offers lower latency and potentially lower energy per bit in selected high-bandwidth topologies. Optical circuit switching, however, still has to integrate with software scheduling, routing, transceivers, packaging, monitoring, serviceability, and data-centre operating practices. The practical challenge is no longer only whether photons can move data quickly, but whether they can be managed predictably inside a commercial AI cluster.
GPU utilisation is a particularly important metric because underused accelerators still consume capital, rack space, power distribution, and cooling capacity. If the network cannot keep processors fed with data, more silicon does not necessarily translate into more inference throughput. Oriole’s claims around idle-time reduction therefore strike directly at the economics of AI infrastructure.
The deployment also sits within Europe’s broader photonics manufacturing and integration push. TNO and ASML’s work on a photonic chip pilot line in Eindhoven is building manufacturing capability around indium phosphide processes, lithography, metrology, and control. Oriole’s system moves further along the chain, where photonic components and architectures must prove their value inside deployed infrastructure.
AMD’s role gives the project a realistic accelerator environment rather than a standalone networking demonstration. ARIA’s Scaling Inference Lab, meanwhile, provides a setting where infrastructure technologies can be assessed around operational constraints instead of narrow benchmark conditions.
AI infrastructure design is becoming less processor-centric as power, cooling, memory, optics, and cluster architecture determine final performance. Oriole’s PRISM installation gives photonic networking a commercial test in one of the hardest places to hide inefficiency: large-scale inference, where every stalled accelerator carries a visible cost.
