PDF 原檔：報告_OCP2026_Meta_800G_AEC大規模部署_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_Meta_800G_AEC大規模部署_20261015_p4.png |
p4 | 真資料圖 | 光模組 vs 被動 DAC vs AEC 比較表：距離（光 10km+／DAC 約 2–3m／AEC <2m）、每端功耗（光 12–15W+／DAC 1–2W／AEC <12W）、成本、112G PAM4 預算、線纜管理 |
報告_OCP2026_Meta_800G_AEC大規模部署_20261015_p5.png |
p5 | 文字卡 | Meta 800G OSFP AEC 規格卡：112G PAM4 lane、32 AWG twinax 約 7.5mm、每端 <12W、2x400G／1x800G／8x100G、CMIS 5.1、1.0–3.2m、Minipack3 對多款 NIC；下方為資格驗證流程 |
報告_OCP2026_Meta_800G_AEC大規模部署_20261015_p11.png |
p11 | 真資料圖 | 系統級輻射 EMI 資格範例：設計餘裕目標 10 dB，各供應商實測約 5.7–9.0 dB；含連接器、編織遮蔽套實物照 |
報告_OCP2026_Meta_800G_AEC大規模部署_20261015_p13.png |
p13 | 真資料圖 | Meta 自建 AEC 長期可靠度（LTR）測試台實物照與參數：樣本 Supplier A 64／B 64／C 32 顆，3 台 MiniPack3，升降溫約 0.7°C/min，目標殼溫約 40／50／70°C，分 36–72 小時、1 週、30 天、90 天階段 |
報告_OCP2026_Meta_800G_AEC大規模部署_20261015_p14.png |
p14 | 真資料圖 | 熱循環下實測殼溫表與 IHS 埠位圖：風扇 45%→最高 64°C、35%→68°C、25%→75°C、20%→82°C（過壓脈衝、非連續）；殼溫超過 70°C 的標準運轉即中止 |
原始內容
p1
Jingyu (Kevin) Yu — Component Quality Engineer, Meta Abhijit Chakravarty — Quality Head - Optics & Interconnects, Meta Large-Scale Deployment of 800Gb/s Active Electrical Cables (AEC) in Meta Data Centers for AI Applications AI OPEN DATA CENTER
p2
Why Short-Reach Interconnects Matter for AI The economics of AI networking are set at the short- reach link.
p3
800G AEC in for X-rack Connectivity
p4
Interconnect Selection: Optics vs DAC vs AEC OPTICS PASSIVE DAC AEC Optical Transceiver + Fiber Passive Copper Cable Active Electrical Cable Reach Up to 10 km+ (Longest reach) Up to ~2–3 m (Short reach only) Up to < 2 m (Short reach) Power (per end) ~12–15 W+ ~1–2 W (Lowest power) < 12 W (Low power) Cost Highest Lowest Low / Medium 112G PAM4 Budget (at required length) Meets easily Cannot meet (Fails budget) Meets with margin (Re-conditions signal) Cable Management More complex (Fiber, patch panels) Simple (Lightweight, flexible) Simple (Similar to DAC) Field Replaceable Yes Yes Yes Fiber Handling Discipline Required (Strict fiber handling) Not required (Copper cable) Not required (Copper cable)
p5
800G AEC in Meta fleet Cable Spec ● 800G OSFP AEC (112 Gb/s PAM4 lanes) ● IHS (switch) to RHS (NIC) connectivity Physical & Power ● 32 AWG twinax, meshed sleeve (~7.5 mm OD) ● Low power design: < 12 W per end Port & Software ● Flex modes: 2x400G, 1x800G, 8x100G ● CMIS 5.1 compliant management Compatibility ● 1.0 m to 3.2 m (intra- /adjacent-rack) ● POR for Minipack3 <-> Multiple NICs How Meta Qualifies AEC: From Lab Validation to Production Scale Component Qualification Product Qualification Single Platform Lab Validation ➔ ➔ ➔ ➔ Cluster Qualification Process Qualification Cross-Platform Multi-Supplier Production Deployment
p6
Five Lessons from deploying Millions of AI Interconnects 01 Component qualification ≠ Cluster qualification 02 Interop failures often originate from firmware rather than SI 03 Manufacturing variation becomes visible only at hyperscale 04 Thermal cycling exposes failure modes hidden by steady-state stress 05 Supplier qualification is a process, not an event
p7
Lesson #1: Component Qualification ≠ Cluster Qualification AI-scale deployment introduces failure modes that are invisible at the component level. Traditional Approach Meta Deployment Approach Component Reliability SI / BER PASS Component Firmware Validation Cross-NIC Interoperability System Traffic Validation (RDMA / NCCL / RoCE) Thermal Cycling Production Cluster Validation PASS in Lab ≠ PASS in AI Cluster ● Component electrical performance ● Environmental reliability ● Connector durability ● SI / BER compliance Sufficient for product qualification, but not for hyperscale AI deployment. ● Multi-vendor interoperability ● CMIS / firmware compatibility ● Production-representative traffic ● Rack-scale thermal behavior ● Manufacturing consistency Example Component Qualification BER clean SI margin pass Environmental qualified Cluster Deployment ✘Link training failed ✘Firmware mismatch ✘Cross-NIC interoperability issue Root Cause CMIS application mode inconsistency Solution Standardized firmware configuration across suppliers Key Takeaway: At hyperscale AI deployment, component qualification is necessary—but not sufficient. System- and cluster-level validation is required before production release.
p8
Lesson #2 Interoperability Lesson: Cross-supplier 01. Failure Bring-Up Stall After reset cycling, NICs never completes datapath bring-up. NIC waits indefinitely, leading to a complete link bring-up failure. 02. Root Cause Firmware Race Condition AEC MCU I2C pre-fetch + firmware latched-status "unerase" logic. Stale reg 0xCA reads 0x1C ("config in progress") instead of 0x11 ("config complete"). 03. The Fix New Firmware Update Assert "config in progress" for < 1 ms to prevent stale reads. Ensure a clean and reliable transition to state 0x11. Key Learning & Test Plan Integration Surfaced only under high-iteration reset cycling added from a prior NIC-AEC investigation — we must carry these critical field learnings directly into the future test plan.
p9
Lesson #3: Manufacturing variation becomes visible only at hyperscale A process that works in one factory or one build does not automatically scale to millions of AI interconnects. 1. Traditional Assumption 2. What Actually Happens at Scale 3. Meta Approach: Process Qualification Factory A Engineering Build PASS Release Traditional Assumption ● Qualified once ● One production line ● One factory ● Small sample size Factory A Factory B Factory C Different equipment Different operators Different material lots Different process windows Variation appears Examples of Variation ● Connector insertion force ● Solder quality ● Glue consistency ● AOI sensitivity ● Test correlation ● Firmware programming Factory Qualification Line Qualification Correlation Study SPC Monitoring MES Traceability On-site Audit Mass Production Meta wants to ensures manufacturing consistency, not just product functionality. Key Takeaway: A qualified design does not guarantee a qualified factory. Production readiness must validate manufacturing consistency—not just product functionality.
p10
Statistical Process Control (SPC): Process Capability Beyond Product Qualification Why do we need SPC? ● Limited sampling ● Snapshot only ● Hidden process drift ● Variation discovered after deployment What process we are watching: Process Monitoring Hot-bar conductor-to-pad SPC Tensile force SPC Conductor width SPC Solder paste volume SPC SMT critical dimensions SPC BER trend (SPC-like) Continuous monitoring The goal shifted from "Can we build it?" to "Can we build it consistently every day?"
p11
Example: System level EMI Compliance 01 System-Level Qualification System-level radiated-emission (RE) qualified: IHS-to-RHS 02 Design Margin Goal 10 dB margin design goal, set for high-cable-count AI racks 03 Iterative Improvements Connector, PCB, and shielding improvements from initial failing states 04 Qualification Results Sources ranged ~5.7 to ~9.0 dB margin; low-margin source improving via silent rolls
p12
Lesson #4 Thermal cycling exposes failure modes hidden by steady-state stress AI clusters experience dynamic thermal conditions that cannot be reproduced by constant-temperature testing alone. Typical Qualification ● High-temperature operating life (HTOL) ● Constant chamber temperature ● Stable workload ● Component-focused reliability check_circleUseful for long-term aging cancelDoes not represent real AI cluster operation Actual AI Cluster Behavior Driven by ● GPU workload ● Fan speed control ● Traffic burst ● Power management Reliability is determined not only by temperature, but by repeated temperature transitions experienced during real AI workloads.
p13
Sample Populations Supplier A 64, Supplier B 64, Supplier C 32 units Hardware & Traffic Setup 3x MiniPack3 (DiagOS, 400G), IHS-to-2xRHS; 800G traffic gen (Spirent / Xena Freya) Thermal Cycling Profile Temperature-cycled, ~0.7 C/min ramp; target case temp ~40 / 50 / 70 C Staged Deployment Pre-test 36-72 hr 1-week 1-month (30d) 3-month (90d) Temp chamber Meta established a deployment methodology
p14
LTR with Thermal Cycling IHS port map (subgroup / lane) ● Case temperature controlled via switch fan PWM ● 45% fan -> max 64 C (cool) ; 35% -> 68 C (nominal) ● 25% fan -> max 75 C (hot) ; 20% -> 82 C (overstress, not continuous) ● Daily 24 h cycling profile + periodic overstress pulses ● Runs halted if measured case temp > 70 C in standard operation ➜Overstress pulses expose marginal behavior without continuous over-temperature stress.
p15
Lesson #5 Supplier qualification is a process, not an event Phase 01 Initial Evaluation 01 Supplier Selection 02 Design Qualification 03 Factory Qualification Phase 02 Production Launch 04 Line Qualification 05 Pilot Build 06 Mass Production Phase 03 Ongoing Lifecycle 07 Production Monitoring 08 Corrective Action 09 Continuous Improvement buildORT & SPC Foundation "Qualification is not the end of supplier management—it is the beginning of production readiness."
p16
Call to Action 01 Deployment Qualification Move beyond component qualification toward deployment qualification. 02 Representative Workloads Validate interoperability, reliability, and manufacturing consistency under production- representative AI workloads. 03 Continuous Lifecycle Treat supplier qualification as a continuous lifecycle— not a one-time milestone. 04 OCP Collaboration Work together through OCP to establish common qualification methodologies for next- generation AI interconnects.
p17
Thank You!