在2026年公开计算项目(OCP)全球峰会上，Meta分享了直接支持我们最新AI模型的网络创新，包括Muse Wonder和Muse Spark，以及我们千兆瓦尺度的Prometheus集群。
要点
- IcePack & IceCube:Meta的第一个液冷网架，有102.4T开关叶片(128×800G,Broadcom Tomahawk 6C和NVIDIA Spectrum-6变体)，运行FBOSS.
- Flattered Planar Fabric(FPF)用于规模化：Meta的多飞机规模化布料，每个GPU与8架网络平面连接。 3级2级Clos,NIC级包喷喷，以及组合主机+开关控制平面。 与NVIDIA共建，与我们的云伙伴共享。
- 规模跨度：BAG，我们的宠物级超级松加入建筑，校园，以及我们千兆瓦级的普罗米修斯集群区域，再加上FBOSS + SAI上的SRv6区域骨干。
- 扩大规模的以太网：Tahan，基于Broadcom Tomahawk5的MTIA 300的以太网升级开关；102.4T硅上每架72个加速器的预览；ESUN从1.0 1.1.
- OPA：由Meta,Microsoft,Alibaba共同牵头的一个新的OPP倡议，实现交换平台层标准化。 v1.0 提交给OCP的规格。
- MetaRoCE：我们为AI网络的RDMA运输，现在去开源。
我们通过许多Meta网络和硬件团队的会谈，分享了对与网络有关的关键举措的见解。 现在每个人都可以参加这些会谈和他们的滑板！
主旨和执行会议
编码混凝土：Gigawat比例尺的工程AI基础设施(视频)
AI改变了基础设施的形态。 Meta的基础设施主管Santosh Janardhan分享了Meta如何设计，建设，运行的规模，需要重新思考网络，设施，以及将网络连接在一起的系统。
发言人：桑托什·贾纳尔丹，基建主管，Meta
在网络上打赌：Meta的"全斯塔克案"，开放AI网络(视频)
AI的进步现在既取决于计算，也取决于网络。 前进的道路是在开放的生态系统中，而不是在专有墙后面，构建一种全面、开放的方法。
发言人：Omar Baldonado,Meta数据中心和AI网络高级主任
突破城墙：AI时代的开放规模交叉(视频)
培训计算复合速度比基础设施规模快。 这一论述说明了为什么需要开放的跨规模联网，以连接各数据中心和区域的人工智能能力。
发言者：Vipul Deokar,VP，核心网络基础设施，Meta;Rakesh Chopra,SVP，硅和系统建筑与工程研究员，思科；Tom Williams,VP，营销，思科
切换平台、 Optics 和互联
Meta 102.4T 以太网开关(视频)
冰块是Meta的102.4T开关，进入我们液冷的冰块架。 Icecube和N-IceCube基于一种2OU，混合液态/空气冷却，PHY无阻的设计，并结合了Broadcom Tomahawk 6C或NVIDIA Spectrum-6 ASICs，具有128x800GE的非阻塞能力。
演讲人：周荣春，网络硬件工程师Meta；网络技术程序管理员雷扎·尼亚兹曼德，Meta
规模集成光学：从51.2TCPO验证到Meta的102.4T(视频)
研究综合光学对机架尺寸的AI——功率，密度，分解和可靠性——为何重要，Meta所学的验证51.2T共包光学，可以承载到102.4T.
发言者：Granthana Rangaswamy，硬件工程师，Meta; Siamak Amiralizadeh，光学工程师，Meta；徐旺，硬件工程师，Meta
超尺度操作线性插孔( 视频)
在AI规模中，每个模块的重时器DSP在功率，热量和成本上都是重复征税，跨越数百万个链接——但移除后将均匀和分级移到主机上。 这段谈话解答了线性光学在舰队规模上是否能够匹配重时可靠性。
发言者：阿明·叶卡尼，光学工程师，Meta；清·王，光学工程师，Meta
在Meta数据中心大规模部署800G AEC(视频)
AI网络的经济学设定在短距离链接。 本会场涵盖Meta大规模部署800Gb/s活动电缆及其背后的组件质量工作。
发言者：Yu Jingyu(凯文)，组件质量工程师，Meta;Abhijit Chakravarty，质量头，Optics & Interconnections,Meta
Byond Pluggables：建设非盈利生态系统以扩大AI
考察构建一个近包装光学(NPO)生态系统需要什么，能够支持AI的缩放，超越可插件模块所能提供的范围。
主讲人：南革，技术探险经理，Meta
FBOSS 和开放平台
开放开关平台架构(OSPA)——统一网络开关平台参考设计(视频).
虽然SAI已经将数据平面软件制成可移植软件，但交换平台层仍保持专有性和零散性，每个平台增加了数月的重复带程工作。 这篇演讲推出 OPA为该层提出了一个统一的，多版本的参考设计，以及已经贡献给OCP的1.0规格。
演讲者：陶任，田芳，和徐汪，Meta；郭汉路，个人；殷雪，Microsoft；黄义元和齐平姚，Alibaba.
NVIDIA SDK + FBOSS：打造生产 - Ready AI网络 Fabrics (video)
NVIDIA的Spectrum-X开关，SDK,SAI的实现，以及模拟器如何与FBOSS,Meta的网络操作系统结合，构建生产AI织物。 会谈涉及CI/CD和测试驱动的工作流程，这些工作流程使快速联合整合成为可能。
演讲者：Parvez Shaikh, Meta;Harish Krishnamoorty, Alan Elder，和Tom Zhu, NVIDIA
打开，AI-Ready FBOSS：从平台带来的解决方案资格(视频)
参观FBOSS软件堆栈——路由，转发，平台服务，以及板面支持——以及一个由配置驱动的方法如何能够更快，更容易地从板面提升到OCP生态系统的合格状态。
发言者：Srikrishna Gopu，软件工程师，Meta;Benoit Sigoure，软件工程师，Nexthop AI
规模推广和规模推广的经验
通过设计实现高性能和可靠性：操作Meta最大的AI集群(视频)
通过FBOSS、司机、通讯库和排程器
发言者：Jose Leitao，生产工程，Meta;Guille Cobo，网络.AI,Meta
扩大华盛顿内外的AI基础设施：Meta与BAG的旅程(视频)
关于后端聚合(BAG)——它的宗旨，架构和性能效益——的概述，从Meta的Prometheus集群的部署中吸取教训。 讨论还涉及平台适配、远程管理、可靠性和能力规划等工程挑战，特别是考虑到迅速发展的AI生态系统。
发言者：Jalpa Patel，技术程序经理，Meta;Hany Morsy，网络工程师，Meta
SRv6 动力反骨网络(视频)
这篇谈话探讨了为什么Meta选择SRv6作为新一层主干线，数据平面和控制平面如何在FBOSS上生产(包括在WAN距离上使用浅缓冲器)，以及几个月后从概念到部署的加速时间表。
发言者：Jasmeet Bagga，软件工程师，Meta;Andrey Golovanov，网络工程师，Meta
搭桥前奏与云：GPU 网络在Meta规模(视频)
考察Meta如何使用云GPU容量，以及Meta带来从其对云的配置中得出什么教训，使其具有高度的性能和对AI工作量的可靠性。
发言者：Adi Gangidi，技术领先，Meta;Carmen Au，软件工程经理，Meta
横跨网络小组(视频)
一个跨行业小组，讨论跨AI集群对路由和光学运输设备的需求。 小组成员比较了超规模供应商、硅供应商和系统供应商的做法。
小组成员：Jeff Rahn, Meta;Tad Hofmeister, Google Cloud;Yawei Yin, Microsoft Azure;Xiguang(亨利) Wu, Broadcom; Vijay Vusirikala, Arista网络(主持人)
缩放和 ESUN
Tahan: MTIA(视频)的开放、以太网为基础的缩放开关
我们介绍Tahan，一个基于以太网的升级开关，用于我们的架式MTIA 300AI系统。 这段话贯穿了Tahan的带宽和耐久性特征以及扩增功能的设置，显示了以太网在扩增和向ESUN发展在未来的扩增设计上的优势。
发言者：Vimal Vasudevan，技术程序经理，Meta;Ashutosh Grewal，软件工程师，Meta
ESUN：为AI缩放时代打开以太网(视频)
去年，我们与社区合作到发布 ESUN：以太网促进规模化网络建设。 本次谈话涵盖了自上次OCP全球峰会以来取得的进展，今年早些时候的1.0规格发布和即将到来的1.1规格。
发言者：Manoj Wadekar，大赦国际网络技术学家，Meta;Pratik Marolia，大赦国际系统建筑师，Microsoft
ESUN生态系统的验证方法(视频)
ESUN生态系统的验证方法，将扩大范围置于更广泛的AI扩展域内——扩大范围、扩大范围、扩大范围、扩大范围、扩大范围、扩大范围以及扩大范围。
发言者：Manoj Wadekar,AI系统技术学家，Meta;Venkat Pulela,CTO,AI基础设施，Keysight
多加速器铜放大式互联(视频)试验方法
在一个铜放大域中，每架数千个连接意味着大约10个坏链接，即使产量为99.9%，而盲友弹匣也没有每个连接的可用性。 在Meta，我们将此视为统计问题，而不是一组孤立的调试问题。 这次谈话分享了我们的测试方法。
发言者：姚荣，系统工程师，AI系统，Meta;Karthik Ganesan,AI/HPC系统性能工程师，Meta
RDMA 运输
MetaRoCE：为AI重新思考RDMA(视频)
网络目前处于AI集群的关键路径，全降级和全降级的性能门由数十万加速器上最慢的节点组成。 这段谈话涵盖了MetaRoCE的运输规格。
发言者：阿尔文德·斯里尼瓦桑，Meta；桑迪普·纳加拉贾，Meta；巴拉克里什南·拉曼，AMD
软件中的MetaRoCE(视频)
Meta正在使其MetaRoCE软件的应用以开源提供，从GitHub(libsoftmetaroce)上的用户空间版本开始，一个基于DPDK的版本即将到来。
发言者：Ashkan Nikravesh，软件工程师，Meta;Yiorgos Nikolaidis，软件工程师，Meta
从2026年OCP全球峰会看一切
你可以在2026年OCP全球峰会网站上观看活动的所有录音谈话。
感谢来自Meta网络和硬件团队的每一位演讲者感谢整个行业的伟大伙伴
图像礼遇Open compute Project(OCP).
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
At the 2026 Open Compute Project (OCP) Global Summit, Meta shared the networking innovations that are directly supporting our latest AI models, including Muse Wonder and Muse Spark, and our gigawatt-scale Prometheus cluster.
Highlights
- IcePack & IceCube: Meta’s first liquid-cooled network rack with 102.4T switch blades (128×800G, Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants), running FBOSS.
- Flattened Planar Fabric (FPF) for scale-out: Meta’s multi-plane scale-out fabric where each GPU connects to eight network planes. 3-tier → 2-tier Clos, NIC-level packet spraying, and a combined host+switch control plane. Built with NVIDIA and shared with our cloud partners.
- Scale-across: BAG, our petabit-scale super-spine joining buildings, campuses, and regions in our gigawatt-scale Prometheus cluster, plus SRv6 regional backbones on FBOSS + SAI.
- Ethernet for scale-up: Tahan, the Ethernet scale-up switch for MTIA 300 based on Broadcom Tomahawk5; a preview of 72 accelerators per rack on 102.4T silicon; and ESUN progressing from 1.0 → 1.1.
- OSPA: a new OCP initiative co-led by Meta, Microsoft, and Alibaba to standardize the switch platform layer. v1.0 spec submitted to OCP.
- MetaRoCE: our RDMA transport for AI networks, now going open source.
We shared insights into key networking-related initiatives through talks from many of the Meta network and hardware teams. Those talks and their slide decks are now available to everyone!
Keynote and Executive Sessions
Concrete to Code: Engineering AI Infrastructure at Gigawatt Scale (video)
AI has changed the shape of infrastructure. Santosh Janardhan, Head of Infrastructure at Meta, shares how Meta is designing, building, and operating at a scale that requires rethinking the network, the facility, and the systems that hold it all together.
Speaker: Santosh Janardhan, Head of Infrastructure, Meta
Betting on the Network: Meta’s Case for Full-Stack, Open AI Networking (video)
AI progress now depends on the network as much as on compute. The path forward is a full-stack, open approach built in the open ecosystem rather than behind proprietary walls.
Speaker: Omar Baldonado, Senior Director, Data Center & AI Networking, Meta
Breaking the Walls: Open Scale-Across for the AI Era (video)
Training compute is compounding faster than infrastructure can scale. This talk lays out why open scale-across networking is needed to connect AI capacity across data centers and regions.
Speakers: Vipul Deokar, VP, Core Network Infrastructure, Meta; Rakesh Chopra, SVP, Silicon & Systems Architecture and Engineering Fellow, Cisco; Tom Williams, VP, Marketing, Cisco
Switch Platforms, Optics, and Interconnect
Meta 102.4T Ethernet Switch (video)
IceCube is Meta’s 102.4T switch that goes into our liquid-cooled Icepack rack. Icecube and N-IceCube are based on a 2OU, hybrid liquid/air-cooled, PHY-less design and incorporate either the Broadcom Tomahawk 6C or NVIDIA Spectrum-6 ASICs with 128x800GE of non-blocking capacity.
Speakers: Rongchun Zhou, Network Hardware Engineer, Meta; Reza Niazmand, Network Technical Program Manager, Meta
Integrated Optics at Scale: From 51.2T CPO Validation to Meta’s 102.4T (video)
A look at why integrated optics matter for rack-scale AI — power, density, disaggregation, and reliability — and what Meta learned validating 51.2T co-packaged optics that could carry over to 102.4T.
Speakers: Granthana Rangaswamy, Hardware Engineer, Meta; Siamak Amiralizadeh, Optical Engineer, Meta; Xu Wang, Hardware Engineer, Meta
Operating Linear Pluggable Optics at Hyperscale (video)
At AI scale, the per-module retimer DSP is a repeated tax in power, heat, and cost across millions of links — but removing it moves equalization and triage onto the host. This talk answers whether linear optics can match retimed reliability at fleet scale.
Speakers: Amin Yekani, Optical Engineer, Meta; Qing Wang, Optical Engineer, Meta
Large-Scale Deployment of 800G AECs in Meta Data Centers (video)
The economics of AI networking are set at the short-reach link. This session covers Meta’s large-scale deployment of 800Gb/s active electrical cables and the component quality work behind it.
Speakers: Jingyu (Kevin) Yu, Component Quality Engineer, Meta; Abhijit Chakravarty, Quality Head, Optics & Interconnects, Meta
Beyond Pluggables: Building the NPO Ecosystem for Scaling AI
A look at what it takes to build a near-package optics (NPO) ecosystem capable of supporting AI scaling beyond what pluggable modules can deliver.
Speaker: Nan Ge, Technical Sourcing Manager, Meta
FBOSS and Open Platforms
Open Switch Platform Architecture (OSPA) — A Unified Network Switch Platform Reference Design (video)
While SAI has made data plane software portable, the switch platform layer remains proprietary and fragmented, adding months of duplicated bring-up work per platform. This talk introduces OSPA, which proposes a unified, multi-vendor reference design for that layer, and the 1.0 specification that has been contributed to OCP.
Speakers: Tao Ren, Tian Fang, and Xu Wang, Meta; Guohan Lu, Individual; Ying Xie, Microsoft; Yiyuan Huang and Zhiping Yao, Alibaba
NVIDIA SDK + FBOSS: Building Production-Ready AI Network Fabrics (video)
How NVIDIA’s Spectrum-X switch, SDK, SAI implementation, and simulator come together with FBOSS, Meta’s network operating system, to build production AI fabrics. The talk covers the CI/CD and test-driven workflow that made rapid joint integration possible.
Speakers: Parvez Shaikh, Meta; Harish Krishnamoorthy, Alan Elder, and Tom Zhu, NVIDIA
Open, AI-Ready FBOSS: From Platform Bring-up to Solution Qualification (video)
A tour of the FBOSS software stack — routing, forwarding, platform services, and board support — and how a config-driven approach can more quickly and easily take a platform from board bring-up to qualified status for the OCP ecosystem.
Speakers: Srikrishna Gopu, Software Engineer, Meta; Benoit Sigoure, Software Engineer, Nexthop AI
Experiences With Scale-Out and Scale-Across
High Performance at Scale and Reliability by Design: Operating Meta’s Largest AI Clusters (video)
What it takes to operate Meta’s largest AI clusters reliably, walking the full training stack from GPUs and scale-up fabrics through FBOSS, drivers, communication libraries, and the scheduler.
Speakers: Jose Leitao, Production Engineering, Meta; Guille Cobo, Network.AI, Meta
Scaling AI Infrastructure Across DC and Beyond: Meta’s Journey with BAG (video)
An overview of Backend Aggregation (BAG) — its purpose, architecture, and performance benefits — with lessons from deployment in Meta’s Prometheus cluster. The talk also covers the engineering challenges of platform fit, distance management, reliability, and capacity planning, especially given the quickly evolving AI ecosystem.
Speakers: Jalpa Patel, Technical Program Manager, Meta; Hany Morsy, Network Engineer, Meta
SRv6 Powered Backbone Networks (video)
This talk delves into why Meta chose SRv6 for a new layer of its backbone, how the data plane and control plane work in production on FBOSS (including using shallow buffers at WAN distances), and the accelerated timeline from concept to deployment in a matter of months.
Speakers: Jasmeet Bagga, Software Engineer, Meta; Andrey Golovanov, Network Engineer, Meta
Bridging On-prem and Cloud: GPU Networking at Meta Scale (video)
A look at how Meta uses cloud GPU capacity, and what lessons Meta brings from its on-premises deployments into clouds to make them highly performant and reliable for AI workloads.
Speakers: Adi Gangidi, Technical Lead, Meta; Carmen Au, Software Engineering Manager, Meta
Scale-Across Networking Panel (video)
A cross-industry panel on what scale-across AI clusters demand from routing and optical transport equipment. Panelists compare approaches across hyperscalers, silicon, and systems vendors.
Panelists: Jeff Rahn, Meta; Tad Hofmeister, Google Cloud; Yawei Yin, Microsoft Azure; Xiguang (Henry) Wu, Broadcom; Vijay Vusirikala, Arista Networks (moderator)
Scale-Up and ESUN
Tahan: An Open, Ethernet-Based Scale-Up Switch for MTIA (video)
We introduce Tahan, an Ethernet-based scale-up switch for our rack-scale MTIA 300 AI system. This talk walks through Tahan’s bandwidth and latency characteristics and scale-up feature set, demonstrating Ethernet’s strengths in scale-up and moving toward ESUN for future scale-up designs.
Speakers: Vimal Vasudevan, Technical Program Manager, Meta; Ashutosh Grewal, Software Engineer, Meta
ESUN: Open Ethernet for the AI Scale-Up Era (video)
We partnered with the community last year to launch ESUN: Ethernet for Scale-up Networking. This talk covers progress since the last OCP Global Summit, with the 1.0 specification released earlier in the year and the upcoming 1.1 specification.
Speakers: Manoj Wadekar, AI Network Technologist, Meta; Pratik Marolia, AI System Architect, Microsoft
Validation Methodologies for the ESUN Ecosystem (video)
A validation methodology for the ESUN ecosystem, situating scale-up within the broader set of AI scaling domains — scale in, up, out, across, edge, and above.
Speakers: Manoj Wadekar, AI Systems Technologist, Meta; Venkat Pullela, CTO, AI Infrastructure, Keysight
Test Methodology for Multi-Accelerator Copper Scale-up Interconnect (video)
In a copper scale-up domain, thousands of connections per rack mean roughly 10 bad links even at 99.9% yield — and blind-mate cartridges leave no per-link serviceability. At Meta, we look at this as a statistical problem rather than a set of isolated debugging problems. This talk shares our test methodology.
Speakers: Yao Rong, Systems Engineer, AI Systems, Meta; Karthik Ganesan, AI/HPC System Performance Engineer, Meta
RDMA Transport
MetaRoCE: Rethinking RDMA for AI (video)
The network is now in the critical path for AI clusters, with all-reduce and all-to-all performance gated by the slowest node across hundreds of thousands of accelerators. This talk covers the MetaRoCE transport specification.
Speakers: Arvind Srinivasan, Meta; Sandeep Nagaraja, Meta; Balakrishnan Raman, AMD
MetaRoCE in Software (video)
Meta is making its MetaRoCE software implementations available in open source, starting with a user-space version on GitHub (libsoftmetaroce) with a DPDK-based version coming soon.
Speakers: Ashkan Nikravesh, Software Engineer, Meta; Yiorgos Nikolaidis, Software Engineer, Meta
Watch Everything From the 2026 OCP Global Summit
You can watch all of the recorded talks from the event on the 2026 OCP Global Summit website.
A huge thank-you to every presenter from our Meta network and hardware teams and to our great partners across the industry who made this possible!
Images courtesy Open Compute Project (OCP).
Auto-fetched and translated by AI Atlas. All rights belong to the original author.