--- 格式版本: 2 标题: "Meta’s custom transport protocol embraces packet chaos to boost AI throughput" 原文链接: "https://www.sdxcentral.com/news/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput/" 发布日期: "2026-08-25" 发布时间校准状态: "found" 发布时间需复核: "否" 发布时间来源: "llm:local:strict_original_body" 发布时间证据: "August 25, 2026 ByBen WodeckiHave your say" 发布时间校准原因: "正文标题下方明确标注作者和日期，格式为典型文章发布时间，且早于发现时间，判定为真实发布时间。" 发布时间校准置信度: "1" 发布时间候选数量: 38 发布时间严格候选数量: 14 发布时间原页读取状态: "source template page reused from URL open" 发布时间未找到原因: "" 发布时间校准时间: "2026-08-27T17:56:08+08:00" 发布时间仲裁状态: "confirmed" 发布时间仲裁尝试次数: 1 发布时间仲裁耗时毫秒: 3945 发现时间: "2026-08-27T17:55:38+08:00" 入库时间: "2026-08-27T09:56:25.313Z" 来源平台: "SDxCentral 搜索" 搜索渠道: "source_template" 搜索词: "site:sdxcentral.com Meta" 匹配关键词: - "latency" - "throughput" - "AI" - "GPU" - "Scale-up" - "performance" - "bandwidth" 相关厂家: - "Meta" - "AMD" 相关专家: [] 内容类型: "网页" 抓取工具: "Jina Reader" 清洗工具: "Jina Reader Markdown + Defuddle/Readability 正文提取" 原始附件: [] 图片摘要: - "✗ ./assets/img-ca27c130.png | ad | Google推广广告，与正文主题无关" - "★ ./assets/img-21a39fb7.webp | diagram | MetaRoCE协议示意图，展示Meta自研传输协议如何应对数据包乱序以提升AI训练吞吐性能。" - "✗ ./assets/img-edd9789a.webp | decorative | 杂志封面推广，与正文无关" - "✗ ./assets/img-f1497174.webp | photo | Google Ironwood TPU相关内容，与Meta传输协议无关的推荐阅读缩略图" - "✗ ./assets/img-13221220.webp | photo | Nvidia Groq LPU相关内容，与本文主题无关的推荐阅读缩略图" - "✗ ./assets/img-98f4b03e.webp | photo | AT&T相关图片，与正文无关的推荐阅读缩略图" - "✗ ./assets/img-bca435ad.webp | screenshot | 未知截图为侧边栏推荐内容，与正文无关" - "✗ ./assets/img-e408dc8f.webp | photo | AI agent概念图，与Meta传输协议主题无关的装饰/推荐图片" AI优质: "是" AI打分: 83 AI分档: "高置信优质" AI质检状态: "通过" AI打分理由: "正文主线是Meta面向AI scale-out网络新开发的MetaRoCE传输协议，并说明其可延伸至机架内scale-up。SDxCentral为专业技术媒体，链接Meta官方工程原文并引用工程师说明。固定知识库仅出现AMD Pensando NIC及其他平台，未覆盖MetaRoCE；本文新增乱序喷洒、端点智能、直接写入最终内存位置等机制，以及AMD Pensando NIC上64节点测试、1%丢包时约86%吞吐、四/八平面和最高4000并发连接的结果，并披露将完整规范开放给OCP及规划其他厂商实现。命中新架构、正式规范路线和生产级深技术通道；尚无量产、客户或规模部署，商业信号较弱。" AI质检模型: "gpt-5.6-sol" AI质检时间: "2026-08-27T17:56:38+08:00" AI主题相关性: 18 AI来源权威性: 13 AI新颖性: 19 AI技术细节: 18 AI商业部署信号: 6 AI完整性: 9 AI评分提示词版本: "v17-精简生产版" AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d" AI评分知识库版本: "knowledge_base_v1-20260819+runtime.7" AI评分知识库SHA256: "2d08daa800e4aaf8f32428770a398a6fddb773ec6de396decbc9d7689bcb65f7" AI评分知识库检索词: "[\"Meta\",\"latency\",\"throughput\",\"site:sdxcentral.com Meta\",\"RAS\",\"GPU\",\"AMD\",\"Intel\",\"URL\",\"SDxCentral\",\"LDN\",\"www.xceleratedcompute.com/london/2026\"]" AI评分知识库命中: "[{\"id\":\"july-correct-0018\",\"title\":\"AMD, Cerebras partner on joint Helios rack-scale AI inference platform\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"Meta\",\"latency\",\"throughput\",\"RAS\",\"GPU\",\"AMD\",\"URL\",\"SDxCentral\"],\"rank\":-20.734547433797633},{\"id\":\"july-correct-0015\",\"title\":\"AMD to join the optical interconnect party with 2027 Instinct GPUs\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"Meta\",\"throughput\",\"GPU\",\"AMD\",\"URL\",\"SDxCentral\"],\"rank\":-17.132485773432286},{\"id\":\"july-correct-0017\",\"title\":\"AMD challenges Nvidia’s networking dominance with Helios racks boasting 50% higher bandwidth\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"latency\",\"throughput\",\"RAS\",\"GPU\",\"AMD\",\"URL\",\"SDxCentral\"],\"rank\":-15.759460306873493},{\"id\":\"july-correct-0016\",\"title\":\"AMD launches Instinct MI400 Series GPUs for AI workloads\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"latency\",\"RAS\",\"GPU\",\"AMD\",\"URL\",\"SDxCentral\"],\"rank\":-14.623845942289933},{\"id\":\"july-correct-0065\",\"title\":\"AMD Pensando™ Vulcano 800 AI NIC: Built to Scale-Out and Across\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"latency\",\"throughput\",\"RAS\",\"GPU\",\"AMD\",\"URL\"],\"rank\":-13.92523542224603}]" AI摘要: "Meta发布自研RDMA传输协议MetaRoCE，面向AI以太网扩展网络，故意让数据包乱序发送以提高吞吐量和利用率。测试显示在1%丢包率下仍保持约86%吞吐量，并计划将完整规范开放给OCP。" AI摘要模型: "ali-deepseek-v4-flash" AI摘要时间: "2026-08-27T18:59:51.668Z" 采集批次: "2026年8月27日17点55分15秒" 采集批次ID: "20260827-175515-082" 去重键: "https://www.sdxcentral.com/news/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput" --- Title: Meta’s custom transport protocol embraces packet chaos to boost AI throughput URL Source: https://www.sdxcentral.com/news/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput/ Markdown Content: This website uses cookies to ensure you get the best experience on our website. [Learn more](https://www.cookiesandyou.com/) [Got it!](https://www.sdxcentral.com/news/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput/) # * Channels * [The AI Channel](https://www.sdxcentral.com/ai/) * [The Networking Channel](https://www.sdxcentral.com/network-sdn/) * [The Cybersecurity Channel](https://www.sdxcentral.com/security/) * [The Distributed Infrastructure Channel](https://www.sdxcentral.com/distributed-infrastructure/) * [The Telecom Channel](https://www.sdxcentral.com/telecom-nfv/) * [The Silicon & Supercomputing Channel](https://www.sdxcentral.com/storage-silicon/) * [The Cloud & Storage Channel](https://www.sdxcentral.com/cloud/) * Media * [News](https://www.sdxcentral.com/news/) * [Analysis](https://www.sdxcentral.com/analysis/) * [Editors' Views](https://www.sdxcentral.com/control-plane/) * [Industry Views](https://www.sdxcentral.com/opinions/) * [Video Interviews](https://www.sdxcentral.com/video-interviews/) * [Broadcasts](https://www.sdxcentral.com/broadcasts/) * [On Demand Broadcasts](https://www.sdxcentral.com/broadcasts/past/) * [Explainers](https://www.sdxcentral.com/sdx-explainers/) * [MarketWatch](https://www.sdxcentral.com/marketwatch/) * [Press Releases](https://www.sdxcentral.com/press-releases/) * [Resources](https://www.sdxcentral.com/resources/) * Live Events * Xcelerated Compute Series * [The Xcelerated Compute Show | LDN 2026](https://www.xceleratedcompute.com/london/2026/) * [The Xcelerated Compute Show | MEX 2026](https://www.xceleratedcompute.com/cancun/2026/) * [The Xcelerated Compute Show | NYC 2027](https://www.xceleratedcompute.com/new-york/2027/) * Digital Infrastructure Series * [Yotta 2026](https://www.yotta-event.com/) * [Login / Create Account](https://www.sdxcentral.com/sdx-en/account/login/?next=%2Fnews%2Fmetas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput%2F) * [](https://www.sdxcentral.com/sdx-en/search/) Search * Show Menu * Channels * Back * [The AI Channel](https://www.sdxcentral.com/ai/) * [The Networking Channel](https://www.sdxcentral.com/network-sdn/) * [The Cybersecurity Channel](https://www.sdxcentral.com/security/) * [The Distributed Infrastructure Channel](https://www.sdxcentral.com/distributed-infrastructure/) * [The Telecom Channel](https://www.sdxcentral.com/telecom-nfv/) * [The Silicon & Supercomputing Channel](https://www.sdxcentral.com/storage-silicon/) * [The Cloud & Storage Channel](https://www.sdxcentral.com/cloud/) * Media * Back * [News](https://www.sdxcentral.com/news/) * [Analysis](https://www.sdxcentral.com/analysis/) * [Editors' Views](https://www.sdxcentral.com/control-plane/) * [Industry Views](https://www.sdxcentral.com/opinions/) * [Video Interviews](https://www.sdxcentral.com/video-interviews/) * [Broadcasts](https://www.sdxcentral.com/broadcasts/) * [On Demand Broadcasts](https://www.sdxcentral.com/broadcasts/past/) * [Explainers](https://www.sdxcentral.com/sdx-explainers/) * [MarketWatch](https://www.sdxcentral.com/marketwatch/) * [Press Releases](https://www.sdxcentral.com/press-releases/) * [Resources](https://www.sdxcentral.com/resources/) * Live Events * Back * Xcelerated Compute Series * [The Xcelerated Compute Show | LDN 2026](https://www.xceleratedcompute.com/london/2026/) * [The Xcelerated Compute Show | MEX 2026](https://www.xceleratedcompute.com/cancun/2026/) * [The Xcelerated Compute Show | NYC 2027](https://www.xceleratedcompute.com/new-york/2027/) * Digital Infrastructure Series * [Yotta 2026](https://www.yotta-event.com/) Close modal ## Search Suggestions [Get the Report: How leading storage architects are solving the AI memory crunch](https://www.sdxcentral.com/resources/the-storage-supplement/) * [Home](https://www.sdxcentral.com/) * [News](https://www.sdxcentral.com/news/) * [The Networking Channel](https://www.sdxcentral.com/network-sdn/) # Meta’s custom transport protocol embraces packet chaos to boost AI throughput MetaRoCE reimagines Ethernet data transfers to fix network latency at AI scale August 25, 2026 By[Ben Wodecki](https://www.sdxcentral.com/profile/ben-wodecki/)[Have your say](https://www.sdxcentral.com/news/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput/#comments) [Facebook](https://www.sdxcentral.com/#facebook)[Twitter](https://www.sdxcentral.com/#twitter)[LinkedIn](https://www.sdxcentral.com/#linkedin)[Reddit](https://www.sdxcentral.com/#reddit)[Email](https://www.sdxcentral.com/#email)[Share](https://www.addtoany.com/share#url=https%3A%2F%2Fwww.sdxcentral.com%2Fnews%2Fmetas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput%2F&title=Meta%E2%80%99s%20custom%20transport%20protocol%20embraces%20packet%20chaos%20to%20boost%20AI%20throughput%20-%20SDxCentral) – Dima Solomin/Unsplash Meta developed its own transport protocol for higher-throughput Ethernet-based scale-out networks running AI workloads. The aptly named [MetaRoCE](https://engineering.fb.com/2026/08/24/networking-traffic/metaroce-rdma-transport-ai-ethernet/) is a remote direct memory access (RDMA) transport designed entirely from scratch. Unlike conventional high-performance protocols, MetaRoCE deliberately sprays packets out of order to achieve mass cross-sectional bandwidth and network utilization. Each write carries its destination in every packet, meaning a send will land correctly even if messages arrive out of order. Meta engineers argued that standard RDMA’s sequential sending of packets stalls network speeds, with receiving network interface card (NIC) hardware being forced to wait until the missing packet arrives to patch the gap. With MetaRoCE, data is written straight to its final memory location as it lands, or as the hyperscaler put it, “the fabric sees packets, but the NIC sees intent.” “Traditional architectures centralize intelligence in the fabric, relying on switches to enforce losslessness and maintain order,” a company blog post explains. “By moving intelligence to the endpoint, MetaRoCE decomposes the network into many fine-grained logical paths, each with its own real-time telemetry.” [](https://media.datacenterdynamics.com/media/images/MetaRoCE-image-1.original.png)![Image 4: MetaRoCE graphic](./assets/img-21a39fb7.webp) MetaRoCE compared to contemporary transport protocols– Meta The Facebook parent tested MetaRoCE on AMD hardware, leveraging the Pensando programmable NICs across a 64-node cluster. Results showed that in packet loss conditions that would typically degrade a protocol like RDMA over converged Ethernet version 2 (RoCEv2), MetaRoCE maintained around 86% throughput at just 1% packet loss. Meta’s home-brewed protocol was found to have consistently maintained higher throughput and lower flow completion times than rival alternatives, even when scaled across four- and eight-plane topologies with up to 4,000 concurrent connections. “By designing for loss from day one and pushing intelligence to the edge, you get a transport that performs better in ideal conditions and degrades gracefully when things go wrong,” Meta engineers wrote. MetaRoCE was teased a few weeks prior at [Advancing AI](https://www.sdxcentral.com/control-plane/5-things-we-learned-from-amds-advancing-ai-2026/), where senior director for data center and AI networking Omar Baldonado and AMD’s Soni Jiandani told [_The Cube_](https://www.youtube.com/watch?v=EUpd1AFZdoE&t=1201s) that the Pensando NIC running the transport would make vast pools of graphic processing unit (GPU) resources more accessible and efficient. Although designed for scale-out networking, Meta said the concept could be applied to scale-up inside the rack to help remove sources of latency, including reorder buffers and priority-based flow control (PFC), as packets are sent out unsequentially by design. While on the distributed computing or scale-across side, Meta engineers claim it could be used to reduce the time between disparate links. Meta confirmed it was opening the full spec to the Open Compute Project (OCP), a move that is set to coincide with the OCP Global Summit in mid-October, and that while it was tested on AMD NICs, additional implementations were planned with “other vendors.” “We’re building this in the open because the challenges ahead benefit from broad industry collaboration. If you’re building NICs, switches, or AI infrastructure, we invite you to join us,” Meta’s blog concludes. Stay ahead of the latest networking channel trends. Subscribe to SDxCentral. [Create an Account to Subscribe Now](https://www.sdxcentral.com/create-your-free-sdx-central-account-today/)