AI News AI资讯 4h ago Updated 2h ago 更新于 2小时前 48

Salience Labs Wants to Scale Up AI with Silicon Photonics Optical Switch Salience Labs希望利用硅光子光开关扩展AI规模

Salience Labs is developing a new optical circuit switch (OCS) based on silicon photonics technology, distinct from existing MEMS and LCoS methods. The startup's OCS aims to enable faster reconfiguration times (under 300 microseconds) compared to traditional MEMS-based switches (milliseconds). Salience Labs' OCS is designed for scale-up networks in high-performance computing clusters, potentially supporting coherent memory clusters across multiple XPUs. The company has partnered with Tower Semic Salience Labs 推出基于全硅光子技术的新型光路交换机(OCS),旨在替代传统 MEMS 和 LCoS 方案,适用于高性能计算集群的规模扩展。 该 OCS 采用相变光电技术(推测),切换速度低于 300 微秒,远快于毫秒级的机械镜或液晶方案,支持动态重构与大规模 TPU/XPU 互联。 产品由两片芯片组成:硅光子 OCS 主芯片 + 专用放大/信号调理芯片,通过 PCB 背对背封装实现低损耗、高可扩展性架构。 公司源自牛津大学与明斯特大学的研究团队,专注于相变光电子学,已与 Tower Semiconductor 合作集成 III-V 激光器与低损氮化硅波导。 愿景是推动数据中心从“

65
Hot 热度
70
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Salience Labs is developing a new optical circuit switch (OCS) based on silicon photonics technology, distinct from existing MEMS and LCoS methods.
  • The startup's OCS aims to enable faster reconfiguration times (under 300 microseconds) compared to traditional MEMS-based switches (milliseconds).
  • Salience Labs' OCS is designed for scale-up networks in high-performance computing clusters, potentially supporting coherent memory clusters across multiple XPUs.
  • The company has partnered with Tower Semiconductor for integrated III-V lasers and low-loss silicon nitride waveguides, indicating the use of phase change technology.
  • Salience Labs' approach addresses the limitations of current OCS technologies by leveraging advanced silicon photonics and amplification chips to reduce optical loss.

Why It Matters

The development of a new OCS technology by Salience Labs could significantly impact the efficiency and scalability of high-performance computing clusters, particularly in AI and machine learning applications. Faster reconfiguration times and reduced optical loss can lead to more efficient data transfer and processing, which are critical for large-scale AI models and distributed computing environments. This innovation may also drive further advancements in silicon photonics and optical networking, potentially disrupting the current market dominated by older technologies.

Technical Details

  • Silicon Photonics Technology: Salience Labs' OCS is built entirely on silicon photonics, which allows for compact and cost-effective integration of optical components on a single chip.
  • Phase Change Switching Mechanism: While the exact switching mechanism is not disclosed, the partnership with Tower Semiconductor suggests the use of phase change materials, which can offer faster and more reliable switching compared to MEMS or LCoS.
  • Amplification and Signal Conditioning Chip: The OCS includes an additional chip for amplification and signal conditioning, addressing the issue of optical loss that typically arises with integrated silicon photonics solutions.
  • Port Scalability: The initial product is a 32-port OCS, with plans to develop versions with 64 and 128 ports, making it suitable for larger-scale deployments in AI and HPC clusters.
  • Back-to-Back PCB Design: The two chips (OCS and amplification) are mounted back-to-back on a PCB card, optimizing space and performance.

Industry Insight

  • Disruption Potential: Salience Labs' OCS technology has the potential to disrupt the current market by offering a more efficient and scalable solution for optical switching in high-performance computing and AI datacenters.
  • Cost and Manufacturing Advantages: By leveraging silicon photonics and integrated manufacturing processes, Salience Labs aims to achieve lower costs per port and better manufacturability, which are crucial for widespread adoption.
  • Future-Proofing: As datacenters increasingly adopt more optical connections, the need for flexible and fast-switching architectures will grow. Salience Labs' OCS is well-positioned to meet these future demands, making it a strategic investment for companies looking to enhance their infrastructure.

TL;DR

  • Salience Labs 推出基于全硅光子技术的新型光路交换机(OCS),旨在替代传统 MEMS 和 LCoS 方案,适用于高性能计算集群的规模扩展。
  • 该 OCS 采用相变光电技术(推测),切换速度低于 300 微秒,远快于毫秒级的机械镜或液晶方案,支持动态重构与大规模 TPU/XPU 互联。
  • 产品由两片芯片组成:硅光子 OCS 主芯片 + 专用放大/信号调理芯片,通过 PCB 背对背封装实现低损耗、高可扩展性架构。
  • 公司源自牛津大学与明斯特大学的研究团队,专注于相变光电子学,已与 Tower Semiconductor 合作集成 III-V 激光器与低损氮化硅波导。
  • 愿景是推动数据中心从“电包交换”向“混合架构”演进,将 OCS 作为 CPO/NPO/XPO 生态中的关键路由层,解决芯片到光纤的数据传输瓶颈。

为什么值得看

本文揭示了 AI 基础设施中一个被低估但至关重要的组件——光路交换机(OCS)的技术跃迁。随着 GPU/TPU 集群规模突破万卡级,传统电气互连已逼近物理极限,而基于硅光子与新材料的新一代 OCS 提供了更高带宽、更低延迟、更可重构的网络拓扑能力,是构建下一代异构超算系统的核心使能技术。

技术解析

  • 核心技术路径:Salience Labs 未公开具体开关机制,但明确排除 MEMS 与 LCoS,结合其与 Tower Semiconductor 的合作(PH18DA III-V 激光器、TPS45PH 氮化硅波导),高度可能采用相变材料(如 GST)实现非挥发性、高速率的光开关控制,属于片上波导切换范畴。
  • 性能优势:相比现有 OCS 方案需数毫秒完成端口重配置,其目标为 <300 微秒,满足 AI 训练中对网络拓扑动态调整的需求,尤其适合弹性调度与容错恢复场景。
  • 双芯片架构设计:系统包含两个功能分离的芯片——一是完全集成的硅光子 OCS 矩阵,二是定制化放大与信号再生芯片(类似铜缆中的 retimer/redriver)。后者用于补偿硅光子集成带来的固有插入损耗,同时通过阵列化制造控制 BOM 成本。
  • 可扩展性与量产导向:当前演示为 32 端口版本,规划支持 64 与 128 端口;强调芯片级集成以实现低成本、高良率生产,符合数据中心规模化部署要求。
  • 供应链协同:依托 Tower Semiconductor 的先进工艺平台(如氮化硅波导、III-V 外延),确保关键无源/有源元件的稳定供应,降低研发风险并加速产品迭代。

行业启示

  • OCS 将成为 AI 数据中心“第二张网”:在 CPO/NPO/XPO 解决“芯片→板卡”短距光互联之后,OCS 将承担“机架间/集群内长距灵活连接”的角色,形成“电包交换 + 光电路交换”的双层混合架构,提升资源利用率与作业灵活性。
  • 硅光子平台正经历从“被动元件”到“主动智能”的转变:过去硅光子主要用于调制器、探测器等静态器件,现在通过引入相变材料与有源放大单元,开始具备可编程、可重构甚至可学习的网络管理能力,标志着光子集成电路进入智能化阶段。
  • 初创企业有机会颠覆成熟市场格局:Google、Nvidia 虽已广泛使用 MEMS-OCS,但其技术根基源于二十年前,存在功耗、体积、速度天花板;以 Salience Labs 为代表的新一代玩家凭借材料创新与系统级优化,有望在高端 HPC/AI 集群中建立新标准,挑战现有供应商生态。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Chip 芯片 Research 科学研究 Deployment 部署