報告_OCP2026_Meta_LPO超大規模營運_20261015

PDF 原檔：報告_OCP2026_Meta_LPO超大規模營運_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

Amin Yekani [Optical Engineer, Meta Platforms] Qing Wang [Optical Engineer, Meta Platforms] Operating Linear Pluggable Optics at Hyperscale: From Link Management to Production Reliability NETWORKING

p2

The problem it solves

At AI scale, the per-module retimer DSP is a repeated tax in power, heat, and cost across millions of links. The challenge - No DSP safety net in the module. The host must own equalization, lost triage capabilities. The question we answer - Can linear optics beat retimed reliability at fleet scale? LPO Benefits and Challenges

p3

This evidence is from 100G/lane LPO in large-scale production, but the lessons carry forward to next-gen linear: 200G/lane LRO/LPO, NPO, and beyond. We shaped it, we productized to it, now we put it to the test. OFC/OCP 2024 Definition Studied feasibility and defined Meta's LPO architecture and spec for the module, informed by the industry MSA and broader ecosystem activity. OCP 2025 Productization How we productize: qualification, host and module calibration, diagnostics, and the path to production. OCP 2026 Reliability at scale Checked LPO reliability at datacenter scale: fleet evidence from live production and a controlled lab.

p4

What changed across the three years. Item OFC/OCP 2024 Definition OCP 2025 Commitment OCP 2026 Data (this talk) Host data rate 2x400G, 8x100G, 1x800G Same Same Media data rate 2x400G, 8x100G, 1x800G Same Same Interop between optics LPO-LPO; retimed-LPO Required: LPO-LPO. In progress: retimed-LPO Limited deployment LPO-LPO Host SerDes setting Same settings per switch vendor/MPN, independent of LPO vendor/MPN (per-port settings allowed) Tx FIR calibration depends on both switch and modules; work in progress on decoupling the two contributions Decoupling now proven, but not used in this deployment (production runs the combined switch + module calibration) Pre-FEC BER target <3E-6 Same Same Fiber link budget Reach = 500 m, IL = 2.5 dB Same Same TP1/a compliance oif2021.405.08 + minor update oif2021.405.11 Same TP2 compliance Retimed lite spec (internal) LPO-specific spec: tightened optical and electrical parameters vs. the retimed lite spec (details internal) Same TP3 compliance Retimed lite spec (internal) Same Same TP4/a compliance oif2021.405.08 + minor update oif2021.405.11 Same 200G/lane Retimed Tx required to be TP2 compliant Same Same Diagnostics New features needed at Driver and TIA Same Same

p5

Setup 1 Live DC production LPO on one hop of a DSF network fabric, LPO on both ends. LPO runs the DTSW-EDSW hop, All the rest are retimed. One question, checked two ways: live production and a controlled lab. Setup 2 Controlled lab isolation ~1k modules in loopback on our MiniPack3 (51T) platform. Strips away DC-specific exposures to isolate module behavior. Link 2x400G FR4, 1x800G mode Placement DTSW EDSW edge Scale 4k modules - 4 months in production Software FBOSS (production network OS) Modules ~1k in loopback Placement MiniPack3 switches Mode 2x400G Software Diag OS stack

p6

Control for age, layer, and mode and the comparison reflects the optics, not the deployment. Comparing Linear Optics to retimed is not a hookup-and-count. Age of the setup New fleets settle as they mature. Compared at equal fleet maturity on the verdict slide. Network layer Each layer takes a different physical path. Link mode How lanes are grouped into links sets both FEC sharing and failure blast radius. Any LPO vs. retimed comparison must therefore use the same link mode.

p7

And the bar only rises. As optics move on-package (NPO, co-packaged), a failed link can no longer be pulled and swapped like a pluggable. Giving up that serviceability only pays off if reliability climbs to match. At AI scale, any interruption can potentially send training back to the last checkpoint. Failure · the old bar Crosses the data center incident threshold. Triaged and service-affecting. Interruption · the new bar Any brief link event, even one below today's triaged threshold. As training tightens what it can tolerate, that threshold has to tighten with it, and interruption rate, not failure rate, is the metric that keeps up.

p8

Same link mode, live production: linear runs ~2× fewer interruptions and ~2.6× fewer service-affecting failures than retimed. - At the module level the two optics are on par, so the gain is architectural: no DSP means one fewer class of failure. - The advantage keeps widening as the fleet matures: ~2× over the full record, more than 8× in the last 60 days. Failure Rate vs. Interruption Rate (Live Production)

p9

Setup

continuous loopback soak at our lab, monitored end to end.

Fan speed is held fixed for the entire run

Stability: 8 lanes, 104 days: optical power holds flat (≤0.25 dB peak-to-peak, ≤0.04 dB per-lane drift) Clean module - Same typical FEC load as retimed, tighter tail, wide margin to codeword loss - Median pre-FEC BER ~1×10−9 - 0 module-attributable failures over 2.96M link-hours More efficient to run (fixed fan speed) - −46% power per module - −17.5°C cooler at that same fan speed Controlled lab environment

p10

• Problem to solve • Deploy LPO at scale without giving up reliability • Build LPO module ecosystem • Try our recipe. Qualify your module/switch and share performance • Where to find us • Amin: ayekani@meta.com • Qing: qingwang@meta.com Call to Action

p11

Thank You!