來源The Next Web (TNW)•較早收集於 2h
英國新創公司聲稱光學交換技術可降低 81% 功耗
💡資料中心功耗若能降低 81%，將重新定義 AI 基礎設施的經濟效益。
⚡ 30 秒速覽
有什麼變化
Oriole Networks 以光學交換器取代電子交換器
為什麼重要
若此技術可擴展，將能顯著降低大規模 AI 訓練叢集的營運成本與環境足跡。
下一步行動
密切關注光學網路硬體的基準測試；這可能是您 GPU 叢集架構下一個重大的效率飛躍。
誰應關注：Developers & AI Engineers
關鍵要點
- •Oriole Networks 以光學交換器取代電子交換器
- •聲稱可降低資料中心網路 81% 的功耗
- •解決 AI 系統中的散熱與頻寬瓶頸
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 17 個來源。
🔑 增強重點摘要
- •Oriole Networks' PRISM platform is being deployed in collaboration with AMD hardware, specifically Instinct GPUs and EPYC CPUs, within the UK's £50 million ARIA Scaling Inference Lab, marking its first commercial use.
- •The optical switching technology is projected to reduce GPU idle time from approximately 60% in conventional systems to less than 1%, which Oriole claims will result in an order-of-magnitude increase in AI inference throughput.
- •Founded in 2023 as a spinout from University College London (UCL), Oriole Networks builds upon more than two decades of optical networking research conducted by co-founder Professor George Zervas and his team.
- •Oriole Networks has secured a total of $34.6 million across two funding rounds, including a $12.6 million seed round and a $22 million Series A round led by Plural.
- •The PRISM solution is designed to support systems with up to one million endpoints with a 1-hop diameter, offering true all-to-all connectivity and handling both deterministic collective communications and non-deterministic dynamic traffic.
🛠️ 技術深入
- Oriole's PRISM platform completely removes electronic packet switches from the network core, replacing them with nanosecond-speed optical circuit switching.
- Data is routed as photons directly from chip to chip, eliminating optical-to-electrical conversions.
- The full-stack solution comprises a high-performance Network Interface Card (providing 800 Gbps bandwidth), an Integrated Photonic Switch and Transceiver (XTR) for on-the-fly wavelength and route selection, and a passive, modular Photonic Router.
- It utilizes nanosecond-level switching and network control across time, wavelength, and space (TDM, WDM, and SDM) to enable rapid circuit configuration, efficiently managing both large and small data transfers.
- The network core features entirely passive interconnects and switches, which means it does not consume power or require cooling.
- PRISM offers true all-to-all connectivity with a 1-hop diameter, capable of scaling to systems with up to one million endpoints.
- The architecture integrates the physical network, AI communication models, and scheduling logic, employing purpose-built algorithms for collective communication strategies optimized for optical circuit-switched networks.
- Its software stack includes xCCL plugins and PCIe drivers designed for seamless integration with existing infrastructure.
- An announced variant, PRISM Ultra, claims a throughput of 50 Exabit per second and 180ns xPU to xPU latency, achieving extensive connectivity with less than 25% of the total system power.
🔮 前景展望基於引用來源的 AI 分析
Oriole Networks' technology will significantly accelerate the adoption of optical networking in AI data centers.
Its demonstrated power efficiency and performance gains, coupled with a commercial deployment with AMD in a UK national lab, provide strong validation for broader industry rollout.
The widespread adoption of photonic networking could lead to a substantial reduction in the environmental footprint of AI infrastructure.
By drastically cutting power consumption and cooling demands, Oriole's solution directly addresses the increasing energy and water usage of large-scale AI systems.
Oriole's vendor-agnostic PRISM platform will foster greater competition and innovation at the GPU layer.
By not being tied to any single chip vendor and running across various accelerator platforms, it offers customers more flexibility and reduces reliance on proprietary networking stacks.
⏳ 時間線
2023
Oriole Networks founded as a spinout from University College London.
2024-03
Raised a Seed funding round of $12.6 million (or £10 million).
2024-10
Raised a $22 million Series A funding round led by Plural, bringing total funding to $34.6 million.
2025-05
Unveiled PRISM (Photonic Routing Infrastructure for Scalable Models), described as the world's first full-photonic network for AI.
2026-03
Announced PRISM Ultra, a one-hop photonic network fabric with 50 Exabit per second throughput.
2026-06
Deployed the world's first large-scale AI system powered by a pure photonic network in collaboration with AMD, as part of the UK's ARIA Scaling Inference Lab, marking its first commercial use.
📎 來源 (17)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： The Next Web (TNW) ↗
每週電子報
每週一封，可隨時退訂。