Oriole Networks, a London-based photonic systems start-up, has announced PRISM — the world’s first pure photonic switched network tailored for data centers, high-performance computing (HPC), and distributed deep learning (DDL) workloads. By tackling network bottlenecks that limit AI training and inference performance, PRISM promises to revolutionize how large-scale AI infrastructure operates with higher throughput, lower latency, and reduced power consumption.
The Network Bottleneck in AI Systems
Despite exponential advances in compute power, real-world AI training and inference are often throttled by limitations in memory bandwidth and network communication. Achieving peak performance is elusive because traditional electronic networks struggle to keep up with the massive data flows required by modern AI workloads.
How PRISM Breaks Through Network Limits
- Nanosecond-Level Switching Across Time, Wavelength, and Space:
PRISM leverages ultra-fast optical circuit switching to reconfigure data paths in nanoseconds. This enables rapid bursts of data transfer optimized for both small and large traffic volumes. - True All-to-All Connectivity:
Unlike conventional networks, PRISM supports port-level all-to-all communication where any node can instantly communicate with any other node, crucial for distributed AI workloads. - Low Power, Passive Core:
The network’s core is entirely passive and requires no power or cooling, reducing operational costs and improving energy efficiency. - Resilient, Fault-Tolerant Architecture:
PRISM’s design eliminates single points of failure by providing multiple paths between nodes, ensuring uninterrupted communication even in the event of hardware faults. - Scalability to One Million Endpoints:
With a 1-hop network diameter, PRISM can scale massively to support highly distributed AI and HPC systems with minimal latency penalties. - AI-Optimized Communication Algorithms:
The system uses purpose-built collective communication strategies to enable schedule-less, contention-free data transfers—critical for keeping GPUs fully utilized during AI training.
PRISM Full-Stack Solution Components
- Software Stack:
xCCL plugins and PCIe drivers integrate PRISM seamlessly into existing collective communication libraries, replacing default transport layers for better performance. - High-Performance NIC:
A compact network interface card delivers 800 Gbps bandwidth compatible with standard server architectures. - XTR Photonic Switch and Transceiver:
Dynamically handles wavelength and route selection, supporting connectivity to up to one million nodes from a single interface. - Photonic Router:
Passive, modular optical routers designed for top-of-rack or central deployments provide dual-path resilience without power consumption.
PRISM by Oriole Networks represents a groundbreaking leap in data center networking, designed specifically to overcome the long-standing bottlenecks in AI infrastructure. By combining ultra-fast photonic switching, energy-efficient passive components, and AI-optimized algorithms, PRISM sets a new standard for scalable, predictable, and power-conscious high-performance computing. This innovation is poised to accelerate distributed AI training and inference, fueling the next generation of AI breakthroughs.
