PDF 原檔：報告_OCP2026_Microsoft_光學ScaleUp架構_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_光學ScaleUp架構_20261015_p7.png |
p7 | 真資料圖 | Blast radius 四種架構對比：共用雷射源→光模組 100%；光板雙雷射 25-50%；CWDM 4 波長 25%；NPC 類 4 路 interleaving 25%（紅色閃電標失效點） |
報告_OCP2026_Microsoft_光學ScaleUp架構_20261015_p8.png |
p8 | 真資料圖 | Proposed Solutions：兩組 Laser Source→Photonic Engine→ASIC 之間交叉接線（綠點標示冗餘與 shuffle 位置），右側為光纖 shuffle 實物（光纖交叉模組與有多個光連接器的測試板） |
報告_OCP2026_Microsoft_光學ScaleUp架構_20261015_p9.png |
p9 | 真資料圖 | Scale-up 拓撲：Accelerator 0…N-1 以光互連連到單層高基數 Switch 0…M-1，上方可再加一層交換器；文字稱光學提供跨機櫃、多排（multi-row）scale-up 所需 reach 與頻寬密度 |
報告_OCP2026_Microsoft_光學ScaleUp架構_20261015_p10.png |
p10 | 真資料圖 | Scale-up 光規格三難題（速度步調、選擇過多、互通）；頻寬 roadmap 表：200Gbps 現在、400Gbps 2027(?)、800Gbps TBD >2030(?)、OCI GEN1 現在；OCI link 示意 GPU–OCI–ELSFP–光纖線束–Switch |
報告_OCP2026_Microsoft_光學ScaleUp架構_20261015_p12.png |
p12 | 真資料圖 | OCI PHY 規格卡片：NRZ 訊號、WDM 8 波長（λ1–λ8，一組標準雷射）、Bi-Di 單光纖；NRZ 眼圖；共同 PHY→OCI MSA→多世代→多供應商生態流程 |
報告_OCP2026_Microsoft_光學ScaleUp架構_20261015_p14.png |
p14 | 文字卡 | 標準組織分工圖：OIF（ELSFP for OCI、CMIS、鏈路可靠度／初始化／互通）、OCP（光托盤與機櫃、form factor、系統可靠度、連接器）、APC（晶圓到系統標準、SiP 製造、CPO/SiP 測試、OSAT） |
原始內容
p1
Adrian Motamedi Interconnect System Architect Fotini Karinou Director of Optical Technologies & Architecture Azure Hardware Systems and Infrastructure (AHSI) Microsoft Optical Scale-Up AI Systems: Architectures, Standards, and Resiliency Strategies OCP SPECIAL FOCUS: PHOTONICS
p2
• Optics make multi-rack, multi-row AI scale-up practical. • System value comes from reach, bandwidth density, efficiency, utilization, and resilience. • Reliability-by-design can reduce failure blast radius and accelerate adoption. • Open standards are needed to align a fast-moving optical ecosystem. OCI-MSA provides a scale-up specification, with OCP playing a key ecosystem role. Key messages
p3
Microsoft Azure Hardware Innovation – Silicon to Systems Silicon Azure Boost DPU Azure Cobalt 100 Azure Cobalt 200 Maia 100 Maia 200 Azure Cobalt 200 Platform Maia 200 Platform Platforms Systems Azure Cobalt 100 Racks Maia 200 Rack Cobalt 200: Azure’s next cloud-native CPU Maia 200: The AI accelerator built for inference Maia 300 Maia 300
p4
Optics make multi-rack scale-up practical REACH Move scale-up beyond a single rack and support multi-row domains. BANDWIDTH DENSITY Pack more bandwidth into accelerator and switch trays. ENERGY EFFICIENCY Reduce pJ/bit pressure as domains grow. UTILIZATION & RESILIENCE Enable better resource utilization with pooled capacity, better resiliency via spares, and new architectures (switch disaggregation). The value proposition is system-level: larger scale-up domains can improve utilization, reduce stranding, and relax rack-boundary constraints. MULTI- RACK SCALE-UP
p5
Optical POD deployment challenges Cost Optical components (lasers, OEs, fibers, shuffles) materially impact system TCO, with uncertainty from evolving form factors, scale-up economics and manufacturing ramp limits. Reliability Millions of links. Each optical link combines many active and passive components (in-tray, in-rack, rack-to-rack), requiring quality, telemetry, operations and failover improvements at scale. System complexity / form factors Form factors’ trade-offs balance cost, reliability and manufacturability as optical connectivity evolves across accelerator, switch and data center integration points. DC operations Optical pods add operational complexity: new infrastructure, fiber management, workload scheduling and RAS strategies to ensure performance and failure mitigation.
p6
Problem statement (Redundancy and reduction of blast radius) • Regardless of the protocol, OBO style, NPO and NPX-style connectors or DR4, OCI, etc… enable high-density electrical-to-electrical/optical interfaces but introduce complex coupled failure domains. • Optical engines share common laser sources and photonic integration. • A single failure (e.g. laser or PIC) can impact multiple channels — many lanes or lambdas can be lost. • Blast radius is far larger than traditional copper links, where failures are isolated and reroutable. • System boundaries often require a single dense connector (NPC/NPO, e.g. CPX MSA, or BOB, etc.), limiting physical separation of redundancy paths. • Industry needs a unified standardized approach to improve resiliency and drive interoperability. Example of need for Reliability Standardization
p7
Redundancy and reduction of blast radius Optical module Or photonic Engine or similar ASIC CWDM ASIC Lambda1 Lambda2 Lambda3 Lambda4 Optical board ASIC Laser Laser NPC or Similar ASIC 100% blast radius In case of optics fail 25-50% blast radius In case of optics fail 25% blast radius In case of optics fail 25% blast radius In case of optics fail. Interleaving. ASIC Laser Source
p8
Proposed Solutions, call for action Photonic Engine ASIC Laser Source Photonic Engine ASIC Areas of redundancies & shuffling Laser Source
p9
Scale-up system topologies Accelerator 0 Accelerator 3 Accelerator N-1
Switch 0 Switch 1 Switch M-1
Accelerator N-4 Switch M-2 Optics provides the reach and bandwidth density required to scale AI systems beyond the rack. Richly connected, multi-Rack and multi row scale-up domains.
---
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
p10
Challenges in AI scale-up interconnect specification SCALING PACE AI systems are evolving faster than “traditional” Ethernet interfaces cycles. OCI MSA gives scale-up optics a multi- generation path at the required pace AI systems need and the required power efficiency. TOO MANY CHOICES The industry needs alignment across optical technologies, rates, wavelengths, signaling, packaging, and link behavior without freezing innovation. ECOSYSTEM INTEROP Scale-up networks need interoperable endpoints and components so suppliers can innovate around a common PHY. OCI-MSA brings industry focus and accelerates AI scale-up innovation through a common multi- generation PHY that enables interoperability and a scalable supplier ecosystem. Bandwidth Spec Timeline 200 Gbps Today 400 Gbps 2027 (?) 800 Gbps TBD, >2030 (?) OCI GEN1 Today Technology - component Package Integration Modulation – rate Modulation - type System Architecture GPU OCI ELSFP Fiber Harness Switch BANDWIDTH ROADMAP WHAT THE SPEC MUST ALIGN OCI LINK: ENDPOINT TO ENDPOINT
p11
Optical Compute Interconnect (OCI) - MSA Founding Members Industry alignment around an optical scale-up specification. Future AI infrastructure built on open standards. Architecture that delivers the efficiency required to scale AI infrastructure. Optical Scale-Up For more information and the OCI-MSA GEN1 Specification please see www.oci-msa.org March 2026 Move Scale-up from In-Rack to Multi-Rack & Multi-Row
p12
OCI Optical PHY Specification for Scale-up INDUSTRY ALIGNMENT Multi-rack scale-up OCI-MSA · March 2026 SIGNALING NRZ Scalable, lowest power OPTICAL SOURCE WDM FIBER EFFICIENCY Bi-Di, one fiber GPU / SWITCH endpoints Physical endpoints COMMON PHY Endpoint interoperability OCI MSA SPEC Focus on tech and KPIs SCALING Path to multi-generation SUPPLIER ECOSYSTEM Multi-vendor, faster adoption A common OCI PHY enables lower power, cost and latency, efficient fiber utilization and multi-generation scaling. λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8 TX RX 8 wavelengths, one standard laser set Reduce cost w/o radix impact Defines optical link layer for AI scale-up
p13
Enabling an Ecosystem for Scale-up Optics Optically Interconnected System Stack Performance, Reliability and Interop Testing Rack and Pod Integration Fiber Shuffles Connectors Server and Switch Platform Integration Packaging Form Factors Optical Engines and Lasers Electrical Interfaces Optical Interfaces Modulators, Photodetectors, Lasers Materials MSAs OIF OCP APC IEEE ITU-T …. Industry Standardization Mechanisms Optical stack complexity is driving urgency for some level of standardization/converge. OIF Compute Optical Interfaces (COI) Whitepaper (requirements for scale-up), OCI MSA (NRZ, WDM, BiDi) Other efforts: Short Reach Optical Interconnects (SROI),Open SiPhotonics for AI Systems (NEW), others Remaining stack still needs ecosystem alignment!
p14
Coordinated scope enables scale-up standards Building on the OCI specification coordinated standards efforts and complementary ownership are needed to align the optical ecosystem across the full technology stack. Example focus areas OIF ELSFP for OCI CMIS improvements Link reliability Link initialization Interoperability OCP Physical optical trays & rack designs Form Factors System reliability Connectors & interfaces Ecosystem requirements APC Wafer-to-Systems Standards SiP manufacturing specifications CPO & SiP test and measurement OSAT-level work OCI MSA momentum
p15
Call to Action: Make OCP the home for scale-up standards JOIN Build on OCI MSA and join adjacent standards work. DEFINE Create complementary OCP workstreams for trays, form factors, reliability. COORDINATE Align OIF, OCP, APC, others, with OCI MSA on complementary scope. DELIVER Turn requirements into open, testable specs. From hyperscaler requirements to open implementation guidance OCP should actively participate in scale-up standards so the ecosystem can move faster together.
p16
Thank You!