At a glance
| Booth | I3000 |
| Country | GB |
| Website | oriolenetworks.com |
Company profile
ORIOLE NETWORKS LTD is a photonic networking company developing technologies for AI/ML and high-performance computing (HPC) that transform data-centre networking, with its work centered on optical communication for large-scale compute infrastructure. Its core innovation, PRISM (Photonic Routing Infrastructure for Scalable Models), is presented as the world's first pure photonic network for AI and as a full-stack architecture built from the ground up for distributed AI workloads. PRISM replaces traditional electronic switching with a deterministic optical fabric designed to deliver predictable, high-performance communication at scale, removing electrical packet-switching bottlenecks so GPU utilization can rise by reducing contention and communication-dependent idle time. ORIOLE NETWORKS LTD has deployed PRISM in the UK ARIA Scaling Inference Lab in collaboration with AMD, using the UK's AI-hardware test bed to demonstrate how next-generation network fabrics can address performance, latency and energy constraints in modern AI infrastructure. The company identifies that ARIA deployment as its first commercial deployment achieved within three years of founding, establishing a commercial milestone for the PRISM architecture. ORIOLE NETWORKS LTD has also been acknowledged in the UK government's £1. 1 billion AI Hardware Plan, placing its photonic-networking activity within a named national AI-hardware initiative. Innovation projects delivered with Innovate UK have resulted in an alpha photonic integrated circuit that enables ultra-fast, multi-dimensional switching. For commercial scaling, the company is developing relationships with organizations in ASIC manufacturing, advanced packaging, testing and back-end processing, and is also building connections with photonics and high-speed design companies for technology partnerships, co-development, design integration and supply-chain collaboration for future PRISM deployment.
Exhibits
PRISM replaces energy-hungry electrical switching with photonic switching and moves data with light alone, with the stated goal of increasing the performance of LLM training and inference while making AI data centres more energy efficient. The network is designed to retain the flexibility of an EPS network while supporting plug-and-grow scaling, large radix with low blast radius and congestion-free operation as the number of connected computing endpoints increases. Its communication design is stated to reduce communication-dependent GPU idle time to less than 1% and to lower power consumption in both the network and compute portions of the cluster. By using less hardware and generating less heat, the architecture reduces the number of failure-prone components and simplifies the overall network structure. Purpose-built algorithms for optical circuit-switched networks enable schedule-free, contention-free data transfers in a small number of steps, reducing latency and keeping GPUs utilized as systems scale. PRISM is designed to handle systems with up to one million endpoints with a 1-hop diameter and provides port-level all-to-all communication so any endpoint can reach any other endpoint. The network has no single point of failure because each node has multiple paths to every other node, allowing communication to continue at reduced capacity if one path or component fails. Entirely passive interconnects and switches allow the network core to operate without dissipating power or requiring cooling, with network control moved to the edge to reduce core complexity. Nanosecond-level switching across time, wavelength and space enables rapid circuit configuration and supports efficient operation for both large and small optical data transfers. PRISM is a full-stack, all-optical fabric designed to connect GPUs at scale while remaining application-independent and using existing infrastructure, so the optical network can be integrated into the current AI/HPC software and hardware stack. The listed AI/HPC applications include Algorithmic Trading, Drug Discovery, Climate Modelling and Natural Language Processing, with Pytorch and TensorFlow as frameworks and NCCL and RCCL as collective libraries used with an Oriole plugin. The integrated software stack also includes CUDA and ROCm GPU runtimes, Libfabric, an Oriole PCIe driver and the Linux kernel; xCCL plugins and PCIe drivers replace default transport layers in collective communication libraries to integrate with PRISM. A pluggable PCIe network-interface and controller card provides 800 Gbps of bandwidth in a compact form factor intended to fit most standard servers. The Integrated Photonic Switch and Transceiver (XTR) handles wavelength and route selection on the fly and is designed to provide connectivity to as many as one million nodes from a single interface. Photonic Routers use a fully passive modular core that operates without power, supports both top-of-rack and central deployments and provides dual-path resilience. The purely passive core is entirely glass with no moving parts, a design intended to improve reliability, minimize loss and eliminate the need for power or cooling in the core. Existing code bases and models are intended to remain unchanged by enabling the PRISM plugin for NCCL or RCCL, allowing the software layer to use the performance characteristics of the fully optical network. The deployment model is plug-and-grow, allowing a customer to start with a single rack and expand during a continuous upgrade cycle, while the existing network can also connect to the fully photonic network for a gradual migration path.
Capabilities and products
- PRISM pure photonic network for AI
- photonic switching replacing electronic switching
- passive photonic routing core requiring no power or cooling
- 1-hop network for up to one million endpoints
- 800 Gbps network bandwidth
- XTR photonic transceiver
- passive modular optical router operating without power
- NCCL/RCCL PRISM plugin software integration
- plug-and-grow photonic network scaling