報告_OCP2026_Meta_CPO從51.2T到102.4T_20261015

PDF 原檔：報告_OCP2026_Meta_CPO從51.2T到102.4T_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

Granthana Rangaswamy, Hardware Engineer, Meta Siamak Amiralizadeh, Optical Engineer, Meta Xu Wang, Hardware Engineer, Meta Integrated Optics at Scale: From 51.2T Validation to Future 102.4T Product OCP SPECIAL FOCUS: PHOTONICS

p2

Integrated Optics Motivation ORv3 HPR MP ORW 2026-Q3 What is next? ≤ 72 Accelerators Cabled Backplane 48 VDC / ± 400 VDC Air / Liquid Cooled IT/Power Rack Single-Wide ≤ 144 Accelerators Cabled Backplane 48 VDC / ± 400 VDC Air / Liquid Cooled IT Rack Double-Wide ≥ 256 Accelerators ? Primarily Liquid Cooled IT Rack Size TBD ? ● Power consumption ● Density ● Disaggregation ● Reliability Siamak Amiralizadeh et al. “Co-packaged Optics Technology Evaluation for Hyperscale Data Center Fabric Switches” invited paper, ECOC 2025 100G/lane

p3

• Meta has collected over 1.8 million system hours of data to assess reliability, availability and serviceability (RAS) of these technologies • A/B test methodology comparing CPO switch systems with pluggable-based Minipack3 platform (control group) • Phase 1 units were deployed in stressed lab temperature condition to gain an initial insight on failure rates and modes • Phase 2 units were manufactured with updated build standards and are intended to give higher-confidence statistics on the interruption rates under normal operation conditions CPO Evaluation Infrastructure Optics Operating Temperature 400G Port Count 400G Device Hours FRO, 2x400G FR4 40 ℃ 1,792 8 M CPO, 128x400G FR4 (Phase 1) 40 ℃ 8,960 50 M CPO, 128x400G FR4 (Phase 2) 15 ℃to 35 ℃ 38,400 175 M FRO: Fully retimed optics

p4

• ELS module, root-caused to a laser drive SMT component, drove most CPO failures. A fix was implemented and verified by design change. • Transient/unknown interruptions were common in FRO but essentially absent in CPO. • Dirty fiber/connector failures stayed consistent across both CPO and FRO and remain a challenge across our deployments. Learnings and Insights I Failure category CPO Phase 1 CPO Phase 2 FRO ELS module 85% 83% NA Fiber/connector 11% 17% 9% Optical engine HW 4% 0% 18% Optics DSP FW NA NA 36% Transient/unknown 0% 0% 37%

p5

Learnings and Insights II • Backend AI network has more stringent reliability requirements and failure thresholds than frontend network • Single uncorrectable codeword (UCW) used as the criteria for “interruption” in the link • Pre-FEC BER is not a reliable predictor of link health as a single metric • Data showed minimal correlation between pre- FEC BER values and link health. • CPO outperformed FRO on failure rate despite worse pre-FEC BER. CPO Phase 1 (8,960 ports) vs FRO Links (1,792 ports) Data is maximum of last 30 minutes before UCW Data is maximum of last 30 minutes before UCW Drew Alduino et al. “A RAS Investigation of Integrated Optical Scale-Up Solutions for AI Interconnects” invited paper submitted to JLT, 2026 Avg Pre-FEC BER ~9e-8 Avg Pre-FEC BER ~1e-9

p6

●Data shows that CPO ports are more reliable than retimed pluggable module ports, potentially due to: ○Reduced number of interfaces and deeper level of integration ○Flexibility in co-design of the system with improved operating condition for key components ○Integration of optics within the platform allows system level test and screening in the manufacturing line ○Reduced human intervention ●No unserviceable failures during the 1.35 million system hours of testing for phase 2 CPO systems ●Field serviceable ELS modules alleviate serviceability and reliability concerns 51.2T CPO RAS Evaluation Optics Assumption ALIR MTBI FRO All interruptions 1.22% 0.72M CPO (Phase 1) All interruptions 0.47% 1.85M CPO (Phase 2) All interruptions 0.2% 4.5M All interruptions except ELS SMT component failures 0.04% 25M Non-serviceable interruptions 0 175M MTBI: Mean time between interruptions ALIR: Annual link interruption rate

p7

Transition to 102.4T with Integrated Optics ●The 51.2T validation proved viability of integrated optics technology for high volume manufacturing ●Next target: 102.4T CPO/NPO product ●Key optical specifications: ○Interoperable with retimed pluggables optics (TP2 and TP3 compliance) ○CMIS 5.3 management interface ○Reach: 500m (spec). Can support up to 2km. ○60% optics power saving compared to retimed pluggables

p8

• CPO and NPO both improve density and shorten the ASIC-to-optics channel, cutting power & cost vs. retimed pluggables • Common considerations: • With power savings established over retimed pluggables, integrated optics’ main hurdle is unknown RAS at scale • Interoperability across multiple vendors, chipsets, and retimed pluggables require TP2/TP3 compliance • Field serviceability with external laser source (ELS) • Reliability is implementation- and vendor-specific Integrated Optics Implementation Trade-offs CPO — Co-Packaged Optics Optical engines integrated directly onto the switch-ASIC package substrate. Advantages ● Shortest electrical path → lowest insertion loss & best signal integrity ● Fewest interfaces / RF reflection points → most scalable to higher lane rates (200G → 400G+) ● System-level RAS partially proven through 51.2T CPO ● Simpler accountability model than NPO Challenges ● Less modular — optical engines are typically soldered to the package and tied to the ASIC vendor's implementation ● Higher optics supply chain risk NPO — Near-Packaged Optics Optical engines mounted near the ASIC linearly driven from the SerDes. Advantages ● Modular interposer approach can foster a broader multi-vendor ecosystem ● Comparable power savings to CPO at 102.4T switch bandwidth (200G/lane) ● Socketed options improve manufacturing flexibility / reworkability. ● Enables integrated and pluggable optics mixed Switch design, which can be required for some network deployments. Challenges ● Longer traces + extra connectors → increased signal integrity and TP2/TP3 compliance risk at scale w/o retimers ● Linear-drive viability degrades at higher lane rates (e.g., 448G/lane); retimers may be needed, eroding the advantage toward pluggable levels. RAS needs to be proven: new mechanics & failure modes

p9

102.4T Integrated Optical Switch Design ●102.4T ASIC with CPO / NPO reduces the electrical lengths and improves signal integrity. ●External laser modules ○Field replaceable laser modules help manage the serviceability risk as the technology matures ○OIF ELSFP compliance to enable open ecosystem and multi-sourcing ●Fibers harness ○Detachable fiber connectors to support manufacturability ○MMC16 front panel optical connectors for high density design

p10

102.4T Integrated Optical Switch Specifications 102.4T CPO/ NPO Switch Form Factor ● 1.5OU, 64 data ports (16x4 MMC16 connectors), 16 ELSFP ports connected to the OEs through Fibre Attach units (FAU). System PCBA’s ● SMB (Switch Main board): Includes the SMB CPLD, switch ASIC, clock, and other peripheral circuits. ● MCB (Main Carrier board): Includes the MCB CPLD, IOB FPGA, COMe Module, and RunBMC card ● LED Board: Features 65 R/G/B tri-color LEDs per PCBA to indicate the status of 64 MMC-16 ports ● ELSFP Board: Equipped with an E1.S connector, QSFP, ELSFP LED. ● IO Board: Contains RJ45 MGMT ports, USB interface, and system status LEDs. ● RunBMC (Baseboard Management Controller) Board: Reused from Minipack 3. ● COMe Module Cooling Optics configuration ● Fit into Meta ORv3-HPRv1 rack in a canister with tool-less ejectors ● 100% liquid cooled due to dust contamination risk (ASIC, CPO/NPO, ELSFP modules)

p11

●Architectural Simplification: Integrated Optics improves high-speed system design by moving complexity directly onto the substrate. ●Board Complexity & Materials: While the high count of power rails necessitates a high layer count board, integrated optics can benefit from the use of more cost- effective, PCB materials. ●Critical Design Challenges: Achieving optimal power integrity and effectively placing power components in relation to optical engines remain significant hurdles that require meticulous mitigation strategies. ●Manufacturing Considerations: Fibre Array Unit pose challenges for reliable cable attachment and routing. 102.4T Integrated Optical Switch Design Factors

p12

102.4T Integrated Optics Switch IO Ports

p13

Switch Blade POC External View

p14

●High-Density Architecture: The 1.5 OU design with integrated optics enables denser switch racks. ●Future-Proof Innovation: Integrated optics enable advanced solutions like optical backplanes. ●Enhanced Signal Integrity : Optical backplanes overcome copper signaling limitations. ●Dust tolerant optical connectors: Provide superior reliability and dust resilience. ●Serviceability: Easier deployment and reduced repair time. Integrated Optical Backplane

p15

• AI networking requires innovative and scalable solutions to keep up with the compute demand in a power-efficient way • Industry-wide effort needed to mature the ecosystem to support future integrated optics technologies • Mature HVM process • Multi-vendor ecosystem • Standardization • Innovations to scale for future generations • 51.2T CPO shows significantly better reliability than pluggables in our evaluation; more data needed across vendors and integrated-optics technologies. • Industry-wide efforts are needed to drive the CPO/NPO technology to a high maturity level. Call to Action

p16

Thank You!