報告_OCP2026_Meta_全端開放AI網路_20261015

PDF 原檔：報告_OCP2026_Meta_全端開放AI網路_20261015_original.pdf

講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，### p<n> 為頁碼；圖表內容見原檔。

圖片清單（已驗證 2026-10-19）

原始內容

p1

Omar Baldonado Senior Director, Data Center & AI Networking Meta Betting on the Network: Meta's Case for Full-Stack, Open AI Networking

p2

The Leap: Muse, an AI agent for everyone

p3

AI Infra has become the Infra

p4

（本頁無可抽取文字，內容在圖中）

p5

Open Full-Stack Networking Innovation

p6

Scale-up / Rack-level Scale-out / Building-level Scale-across / Beyond building Shared hardware & software The Scope of Network

p7

Scale-up / Rack-level Scale-out / Building-level Scale-across / Beyond building Shared hardware & software Scale-out

p8

102.4T Scale-Out: IcePack & IceCube / N-IceCube ●IcePack: our first liquid-cooled network rack for high power density & thermal efficiency (based on ORv3-HPRv1) ●IceCube / N-IceCube: our 102.4T switch, with Broadcom TH6 / NVIDIA Spectrum-6, running FBOSS + SAI software (systems from Celestica and Accton) IcePack

p9

Flattened Planar Fabric: Our next scale-out fabric ●“Multi-plane fabric” that connects accelerators to multiple fabric planes (removing top-of-rack) ●2-tier => fewer switches/optics & lower latency ●NIC-level load balancing ●Combined switch & host control plane ●Close partnership with NVIDIA

p10

Working with NVIDIA & clouds on multi-plane ●Leveraging Meta’s unique position as both hyperscaler with 1P infra and cloud customer ●Saw heterogeneous multi-plane implementations in clouds, e.g., load balancing, topologies, failure domains ●Sharing design & operational learnings from Meta’s and NVIDIA’s FPF development ⇒ Benefit joint cloud partners with higher reliability and performance

p11

Scale-up / Rack-level Scale-out / Building-level Scale-across / Beyond building Shared hardware & software Scale-across

p12

Prometheus: 1+ GW site (partial zoom-in)

p13

Prometheus Inter-Site Connectivity: BAG (Backend Aggregation) Network

p14

Making network perform for large Muse models 1. Keep availability high: interruptions slow down jobs, and frequent interruptions at this scale (millions of components across all scale- up/out/across networks) 1. Implement systems to handle faults: Parallelism-Aware Fault Tolerance (PAFT) system isolates restarts to smaller clusters; optimized checkpointing 1. Iterate quickly on co-design of models & algorithms: reduce exposed communications through tight measurement ⇒ optimization loop

p15

Building regional backbones (RBB)

p16

Using SRv6 on FBOSS + SAI for RBB ●Implemented Segment Routing over IPv6 (SRv6) in FBOSS using SAI ●Motivated by speed, control, and flexibility ●Leveraged shallow buffer switches, starting with Cisco Silicon One for initial production ●Design to production in 5 months, followed by 3 months of long-haul and traffic engineering/SDN controller production

p17

Scale-up / Rack-level Scale-out / Building-level Scale-across / Beyond building Shared hardware & software Scale-up

p18

Ethernet for Scale-Up in MTIA: Tahan ●Tahan provided Ethernet-based scale-up for MTIA 300 (Meta Training & Inference Accelerators) ●Used Broadcom TH5 51.2T ASIC ●Leveraged FBOSS as NOS ●Looking ahead to Broadcom TH6 design ●Subset of features to optimize for scale-up, prepping for ESUN-compliant ASICs

p19

OCP Ethernet for Scale-Up Networking (ESUN) progress ●Multi-vendor collaboration in OCP - thank you! ●1.0 Spec published in March ●1.1 Spec working on continued latency reductions

p20

Scale-up / Rack-level Scale-out / Building-level Scale-across / Beyond building Shared hardware & software Shared HW & SW

p21

Open Switch Platform Architecture ●Problem: slow onboarding of new platforms, duplicate engineering, lifecycle cost ●Solution: vendor-neutral and NOS-neutral platform- layer specification across power, sensors, thermals, fan, inventory, out-of-band interfaces, firmware, troubleshooting ●Community-driven: started by Meta, Microsoft, Alibaba, now a dozen OCP members contributing! ●Status: Specification 1.0 in review

p22

FBOSS: more open, config-driven, & AI-friendly ●Similar problem: reduce time to onboard a new platform to FBOSS itself ●Solution: make FBOSS more config-driven, more in open- source (e.g., routing protocols), and AI-friendly for coding and operations - partnership with Nexthop AI ●Wins: ○Multiple platforms now coming up faster without Meta help ⇒ faster intro into Meta production ○AI skills in open-source usable with any coding agent

p23

MetaRoCE: Meta’s RDMA transport for AI ●Protocol announced in August ●Addresses RoCEv2 limitations with: ○RDMA transport on best-effort Ethernet ○Multi-path handling inside a queue-pair ○Out-of-order tracking & stream multiplexing ●Already supported by AMD in Pensando NICs ○Working with NVIDIA

p24

MetaRoCE: now in open-source! Today at OCP: Meta is contributing two implementations of MetaRoCE for hardware / protocol developers: ○libsoftmetaroce: user-space implementation ○dpdkmetaroce: performance with kernel bypass

p25

Scale-up / Rack-level Scale-out / Building-level Scale-across / Beyond building Shared hardware & software Congrats to teams, thanks to partners

p26

The future is for everyone - let’s build it together!

p27

（本頁無可抽取文字，內容在圖中）