PDF 原檔：報告_OCP2026_Microsoft_Meta_ESUN乙太網路ScaleUp_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）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_Microsoft_Meta_ESUN乙太網路ScaleUp_20261015_p4.png |
p4 | 真資料圖 | scale-up 網路位置示意（CPU／Accel 機櫃內綠色 Scale-Up Network，與 Front End、Scale-out 分開）與 compute tray—low latency switch tray 全互連圖；文字：GPU 數 8→72→576→1K，每 GPU 頻寬 450→1200→1800 GBps |
報告_OCP2026_Microsoft_Meta_ESUN乙太網路ScaleUp_20261015_p7.png |
p7 | 真資料圖 | ESUN 1.0 壓縮封包格式：以 4B EH header 取代 28~48B IP header，E2E transport header／payload／E2E CRC／FCS 保留；文字稱省 24~44 bytes |
報告_OCP2026_Microsoft_Meta_ESUN乙太網路ScaleUp_20261015_p8.png |
p8 | 真資料圖 | ESUN header 欄位：Rev 2b、F、EH-CoS 3b、EH-ECN 2b、Flow label 16b、TTL 4b、UD 2b、RSVD 2b |
報告_OCP2026_Microsoft_Meta_ESUN乙太網路ScaleUp_20261015_p10.png |
p10 | 文字卡 | ESUN 1.1 進行中項目：Switch Profiles（Marvell Meg Lin／Cisco Ramesh Sivakolundu）、UFH 0.9、IPG Reduction 0.9、1.0 勘誤；後續候選：低延遲 FEC、Reduced Lane、In-Network Collectives |
報告_OCP2026_Microsoft_Meta_ESUN乙太網路ScaleUp_20261015_p13.png |
p13 | 文字卡 | ESUN 時程：2025-10 公布、2025-12 成立工作組、2026-03 ESUN 1.0 發布、2026-10~11 ESUN 1.1 目標（依 UEC 更新規格）、2.0 時程 TBD |
原始內容
p1
Manoj Wadekar AI Network Technologist, Meta Pratik Marolia AI System Architect, Microsoft ESUN: Open Ethernet for the AI Scale-Up Era NETWORKING
p2
AI Clusters Growing
p3
AI Networks: Trillions of wires, Multiple networks
p4
Scale-Up Networks : Focus of ESUN • Number of GPUs: 8 🡪 72 🡪 576 🡪 1K • Bandwidth per GPU (GBps): 450 🡪 1200 🡪 1800 🡪 .. • ESUN was formed to create an open solution
p5
• Champions Ethernet as an Open interconnect for Scale-Up by specifically identifying critical features. • Leverage strong Ethernet eco-system of NOS, operational and performance tools • Establishes clear Operator/End User requirements ESUN Progress • Mission: Transitioning scale-up AI interconnects from proprietary protocols to open Ethernet standards
p6
ESUN Technical Focus • Focus On Fabric • For wider adoption while allowing implementation flexibility for GPUs • Standards Coordination: • Leverages Standards work in IEEE, UEC ESUN 1.0 Spec Released March 26!
p7
○Traditional IPv4 was designed for dynamic, routable, and large-scale networks ○Scale-up network is fixed topology. ESUN header replaces IPv4/UDP to improve header efficiency - gives 24 ~ 44 Bytes savings ○Pros: higher bandwidth efficiency for smaller messages ESUN 1.0: Compact Framing
p8
• ESUN header drops IP header, but carries forward useful features to support to build a multi-hop fabric • EH-ECN: provides congestion feedback • EH-COS: identifies differentiated traffic classes • Flow label: fabrics can use this to achieve load balancing • TTL: detects unintended loops • UD: endpoint defined bits, opaque to the fabric ESUN 1.0: Header ESUN header
p9
○Link level retry (LLR) provides localized, faster recovery & lowers endpoint buffering ○CBFC enables tighter buffer sizing, not packet drops due to buffer overflow ○Pros: • Enables simpler transports, simpler endpoints • Shallow buffer switches Note: LLR and CBFC are defined by UEC ESUN 1.0: Reliable Network Fabric
p10
ESUN 1.1 - New Work Items in progress ESUN 1.1 Specification Future Release Candidates "Switch Profiles" Meg Lin (Marvell) / Ramesh Sivakolundu (Cisco) Items Being Worked for ESUN 1.1 ● Optimized Headers (UFH): UFH 0.9 to be contributed to UEC 1.1 — WIP ● IPG Optimization: IPG Reduction 0.9 to be contributed to UEC 1.1 — WIP ESUN 1.0 Errata Low Latency FEC Modes Depends on UEC WG timeline Reduced Lane Mode Depends on UEC WG timeline In-Network Collectives (INC) TBD. Started educational presentations from UEC
p11
ESUN 1.1: UFH 0.9 Transport ESUN Header ESUN Header UFH Header
p12
• IEEE Standard requires average of 12 bytes inter-packet-gap • Also, must pad out after terminate to 8B boundary with idles • Minimum IPG will be 1-8 bytes depending on frame size • Deficit idle count may force additional 8B blocks of idle • Proposed option to only require one IPG byte (terminate symbol) • Retain frame always starting on 8B boundary • Resulting in 1-8 bytes IPG to reach frame boundary • Defined in upcoming UltraEthernet specification ESUN 1.1: IPG Idle Reduction
p13
Roadmap & Working Group Priorities October 2025 ESUN Announcement December 2025 ESUN Group Formation March 2026 ESUN 1.0 Spec Published Oct - Nov 2026* (WIP, UEC Updated Spec dependent) ESUN 1.1 Spec Target TBD ESUN 2.0 Spec Target
p14
• Problem to solve • Jointly advance Ethernet fabrics optimized for scale-up networking • How to get involved in the project/sub-project community • Project Wiki: https://www.opencompute.org/w/index.php?title=Networking/ESUN • Sign up for mailing list: OCP ESUN Mailing List • Join bi-weekly call: OCP Networking Project Calendar Call to Action
p15
Thank You!