PDF 原檔：報告_OCP2026_Microsoft_Arista_XPO系統評估_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_Arista_XPO系統評估_20261015_p5.png |
p5 | 文字卡 | Scale-up 光互連技術三階段評估流程（零件就緒、系統設計就緒、部署就緒），註明每世代 2-3 年 |
報告_OCP2026_Microsoft_Arista_XPO系統評估_20261015_p6.png |
p6 | 真資料圖 | Intro to XPO 要點（204.8T/1OU@200G、409.6T/1OU@400G、液冷至 500W/模組）＋XPO 模組爆炸圖＋MSA 核心光模組成員 logo 牆（20 家） |
報告_OCP2026_Microsoft_Arista_XPO系統評估_20261015_p7.png |
p7 | 真資料圖 | 加速器側光模組外形比較表：Pluggable（XPO/OSFP）、OBO、NPO、CPO MCM、CPO Interposer，逐項列服務性、爆炸半徑、標準化；結論列「推論頻寬 1-2 世代不需更深整合」 |
報告_OCP2026_Microsoft_Arista_XPO系統評估_20261015_p9.png |
p9 | 真資料圖 | Arista 製作的 212Gbps SerDes 模擬測試箱：4 顆 Broadcom Agera3 retimer、NPC 銅 flyover 纜線 300/600/900mm、整合液冷；含插損預算表（總計 15.42dB）與內部實拍 |
報告_OCP2026_Microsoft_Arista_XPO系統評估_20261015_p12.png |
p12 | 真資料圖 | FRO-FRO XPO 模組 BER/SNR 圖：交換器側與加速器側 Host BER 對 flyover 纜線長度（14/18/22dB），0dB 與 3dB 光衰減兩條線；交換器側 BER 貼在 floor 10^-13、加速器側 22dB 時略升至約 10^-12，下排為 Host SNR 約 23-25dB |
報告_OCP2026_Microsoft_Arista_XPO系統評估_20261015_p16.png |
p16 | 真資料圖 | LPO-FRO XPO 模組 BER/SNR 圖：交換器側 BER 約 10^-10 量級、加速器側圖上約 10^-8～10^-9（投影片文字稱在 10^-7 範圍內），交換器側 SNR 約 22-23dB（加速器側 SNR 省略）；橫軸 15/19/23dB（300/600/900mm） |
原始內容
p1
Daniel Mohaghegh Hardware Engineer, Microsoft Sunil Priyadarshi Senior Director, Arista System-Level Evaluation of XPO for High- Density Optical Interconnects in AI Infrastructure OCP SPECIAL FOCUS: PHOTONICS 1
p2
• XPO High-Density Pluggable Form Factor, as applied to AI Scale-Up • Detailed XPO platform Test Data • Q&A Agenda 2
p3
XPO High-Density Pluggable Form Factor, as applied to AI Scale-Up 3
p4
AI Scale-up System Topologies Richly connected, Multi-Rack and Multi Row Scale-up Domains Several topology variations are possible depending on target AI workload performance and cost Reach, Bandwidth and Connectivity enabled by optical interconnects in Accelerator trays High Radix switches in a Single Tier enable uniform communication domain Additional tier of switches for further scaling Example Topology 4
p5
Scale-up Interconnect Technology Evaluation: Towards at-scale deployment Ph 1: Ingredient Readiness Prioritize technologies Integration complexity Link performance and power Ph 2: System Design Readiness End-to-end link performance Optical device reliability Accelerator/Switch/Rack design Fiber management Ecosystem / Supply chain Ph 3: Deployment Readiness DC installation & operations RAS Telemetry Firmware and Manageability Control plane software Data plane software Increasing scope, down-selection of technologies, 2-3 year span per generation 5
p6
1. Highest Density Pluggable Optics form factor in the Industry 2. 204.8T per 1OU with 200G/lane, 409.6T per 1OU with 400G/lane 3. Supports all existing and emerging optics and interconnect standards 4. Liquid cooling up to 500W per module (for 25.6T coherent-lite) 5. Much lower failure rates with fewer components and lower temperatures 6. Vibrant XPO MSA ecosystem supported by all optics module vendors Intro to XPO 6
p7
SoC Board: Near-Packaged Optics (NPO) Main PCB / Board Organic Substrate Interposer SoC IOD IOD NPO ELS On-Board Optics (OBO) Separate (Off-) Board: OBO Organic Substrate Interposer SoC IOD IOD Main PCB / Board IO Board OBO NPC ELS Cu flyover CPO Multi-chip Module (MCM) Main PCB / Board Organic Substrate Interposer SoC IOD IOD CPO ELS Co-Packaged Optics (CPO) Pluggable Optics XPO or OSFP Main PCB / Board Organic Substrate Interposer SoC IOD IOD NPC XPO or OSFP with ILS IO Board Cu flyover CPO Interposer Main PCB / Board Organic Substrate Interposer SoC IOD IOD PIC ELS FAU Power LPO, TRO, FRO Design Scope Tray Serviceability Module-level Blast Radius Module Standards Yes LPO, TRO, FRO Tray Tray-level Module No Linear optics SoC Board Tray-level Module(with CPX) Open CPX Low SoC Package Tray-level Module & Accelerator No Very low SoC Interposer Tray-level Module & Accelerator No Serviceability: The granularity at which a failed optical module can be serviced. Blast Radius: Cost of replacement (module: only module has to be replaced; accelerator: module and associated accelerator has to be replaced) Above blast radius definition is in terms of cost of hardware replacement. It is not equivalent AIR or Blast Radius on workload(s) Industry Directions on Accelerator-Side Optical Module Form Factors Inference workload bandwidth per SoC is not requiring deeper integration for 1-2 generations XPO provides Serviceability, Low Blast Radius and Flexibility as a broad industry standard Accelerator board constraints may encourage exploration of XPO for scale-up optics 7
p8
Detailed XPO platform Test Data 8
p9
An emulation box with 212 Gbps SerDes was designed by Arista to exercise the 64 channels on the XPO modules. This emulation platform featured the below characteristics. - x4 Agera3 Broadcom Retimers • x16 lanes 212 Gbps each - NPC Copper Flyover Cables • 300/600/900 mm long • 32 AWG Twinax: Amphenol XPO to Ultrapass-XR - Integrated Liquid Cooling XPO Module Test Setup 9
p10
XPO Module Test Setup: End-to-End Connections For each module version, two systems were connected, a switch on the left with fixed cable and an accelerator on the right with three flyover cable lengths. Host BER/SNR and module RX power data are collected across: • Different flyover cable length • W/Wo 3dB optical attenuators 10
p11
XPO Module Test Setup: Accel Flyover Cable Lengths The switch side flyover cable length was fixed at 300mm (15dB bump-to-bump loss). 15dB 19dB 23dB 300mm 600mm 900mm 11
p12
●Switch-side BER remains independent of accelerator flyover cable length ●Accelerator-side BER shows minor degradation with loss ●FRO link demonstrates negligible performance differences between 0dB and 3dB loss scenarios. FRO-FRO XPO Modules: BER/SNR 12
p13
FRO-FRO XPO Modules: Rx Power • Currently, the optical power of one module is high. As module design continues to improve, power will become more standardized. 13
p14
LRO-LRO XPO Modules: BER/SNR 14 ●Switch-side BER remains independent of accelerator flyover cable length ●Accelerator-side BER shows minor degradation with loss ●Due to the lack of Rx-side DSP, optical performance is sensitive to optical loss
p15
LRO-LRO XPO Modules: Rx Power 15
p16
LPO-FRO XPO Modules: BER/SNR 16 • Switch-side BER floor stays within 10⁻¹⁰ • Accelerator-side BER remains within acceptable 10⁻⁷ range, primarily constrained by LPO Tx performance • Accelerator SNR omitted as it is not meaningful • Link performance expected to improve with standardized power, drivers, and CTLE settings
p17
LPO-FRO XPO Modules: Rx Power 17
p18
BER Comparison versus modules pairing 18 • FRO: Provides the strongest BER margin and robust performance across the tested conditions. • LRO: Good performance on both switch and accelerator sides, with greater sensitivity to optical attenuation. • LPO: Demonstrates good switch-side performance; further Tx optimization and tuning can improve link margin.
p19
Next Steps From Demonstration to Deployment • Expand testing to additional XPO module vendors and optical architectures • Validate interoperability across FRO, LRO and LPO module combinations • Optimize LPO Tx performance and establish interoperable tuning guidelines • Progress from emulation-platform testing to switch and accelerator system-level validation 19
p20
Call to Action Build an Open and Interoperable XPO Ecosystem • Drive multi-vendor interoperability across XPO implementations. • Align host, module, and optical interfaces to enable plug-and-play interoperability. • Establish common system-level test methodologies and performance metrics. • Foster collaboration among system, silicon, and optics vendors to accelerate XPO adoption in AI scale-up networks. • Enable new applications and innovations leveraging the XPO platform, including XPO ACC/AEC, OCI optics, coherent optics, external light sources, and emerging optical architectures. 20