報告_OCP2026_Celestica_ScaleUp互連轉折點_20261015

PDF 原檔：報告_OCP2026_Celestica_ScaleUp互連轉折點_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

Akarsh Bhatnagar Product Line Manager, Celestica The Interconnect Inflection Point: Evolving Intra- and Multi-Rack Scale-Up AI Networks OCP SPECIAL FOCUS: PHOTONICS

p2

1. System-Level Convergence 2. XPO: The Interconnect “Swiss Army Knife” 3. Closing The Critical Reach Gaps 4. Promising Technologies 5. Interconnect Blueprint 6. CALL TO ACTION: Get involved with the open standards Agenda

p3

• As AI/ML clusters expand from single-rack to multi-rack architectures, the scale-up interconnect has become the primary design bottleneck. • Power and cooling • Front-panel thermal limits and excessive power consumption caps module packaging density. • Physical reach limits • Sub-2m copper reach; can no longer reliably span a full rack • Interconnect latency • Reconditioning degraded signals with DSPs Interconnect Bottlenecks

p4

• Industry is converging on complementary open specifications spanning the full path from front panel to silicon packaging, preserving multi-vendor ecosystems System-Level Convergence MSAs Pluggable Density Pivot On-Board & Packaging XPO (eXtra-dense Pluggable Optics) Next-generation form factor redefines front-panel interconnect density: 12.8T Module Capacity: Enables up to 204.8T total bandwidth density within a single 1RU switch chassis. Liquid Cooling Integration: Direct blind-mate cold- plate cooling engineered for modules requiring up to 400W+ per slot. Open CPX & OCI System-level standardization targeting near-die and silicon- centric connectivity Open CPX MSA: Standardizes socketed, pluggable optical & copper engines for Near-Packaged Optics (NPO), creating an open multi-vendor ecosystem. OCI MSA (Optical Compute Interconnect): Silicon-centric scale-up paradigm taking slow-wide approach using DWDM supporting bidi links.

p5

• One footprint, many engines: Swap engines as per requirements, without a redesign • Retains field-serviceability and hot-swappability while achieving CPO-class density DAC Direct Attach Copper Shortest reach, lowest power, lowest cost - in- rack only XPO: The Interconnect “Swiss Army Knife” AEC Active Electrical Cable DSP/retimer-based; extends copper reach for intra-rack links ACC Active Copper Cable Linear-equalizer copper; ~2.5W, short reach, minimal power XPO Shared mechanical & thermal envelope, matched to reach and cluster tier, eliminating the redesign cost of switching interconnect types LPO / LRO Linear (Re)drive Optics DSP-free or lightly retimed optics; lower power, shorter reach FRO Fully Retimed Optics Full signal reconditioning for longer reach, higher robustness ZR / Coherent-lite Coherent/ Coherent-lite Engines long-reach coherent optics for scale-across fabrics

p6

Closing the Critical Reach Gaps • No single electrical or optical engine covers the full reach continuum a rack-scale AI cluster needs. • Why it matters: Prevents architectural & reach gaps as AI clusters expand across racks. Passive DAC ≤ 1.0m (224G) Strict channel loss cliff INTRA-RACK ACC 1.2m – 2m Analog equalizer cable AEC 2m – 5m Retimed active copper LPO Optics 2m – 10m+ (No DSP) Ultra-low latency scale-up LRO/ FRO (Retimed) 10m – 500m Coherent – ZR , Lite 500m – 10km+ Scale-Out/ Across, DCI SCALE-OUT / Across Scale-up/ Scale-Out Power Consumption

p7

Promising Technologies MicroLED Optical Engines • A “wide-and-slow” architecture utilizing arrayed MicroLED emitters for massively parallel optical channels. • Can be integrated directly onto silicon, operating at sub-picojoule per bit energy efficiencies. Active Radio Cables (ARCs) • Employs RF/microwave guided-wave media as an alternative to copper and lasers. • Standard pluggable connectors & formfactor. • Achieves ultra-lightweight, high-frequency intra-rack connectivity with lower loss. • Expected reach 2m-10m

p8

Interconnect Blueprint: Intra- to Multi-Rack • No single technology wins across the whole fabric. • The winning architecture balances density, serviceability, and thermal efficiency differently at each tier. • Liquid cooling Next-generation switch integrate dripless quick-disconnect liquid manifolds to handle module thermal loads. Intra-Rack (Scale-Up) DAC / ACC / AEC Emerging : ARC/ MicroLED Density & power dominate; serviceability is secondary at short, fixed reach. Rack-to-Rack (Scale-Up/Out) CPO / NPO / Open CPX / XPO Thermal efficiency dominates; silicon-adjacent optics cut electrical loss. Row / Multi-Rack (Scale-Out) XPO pluggables (LPO/LRO/FRO) Serviceability & density balance; liquid-cooled pluggables preserve field swap. Row / DC (Scale-Across) Coherent - Lite, ZR/ZR+ Reach dominates; coherent- lite and ZR/ZR+ optics extend the fabric.

p9

Production Adoption Requires Proof at Scale Breaking the Validation Deadlock Proof at Scale Requires Real Deployments Phase 1: Managed Risk XPO & Socketed NPO Liquid-cooled. Swap-ability/ serviceability isolates optical failure risk. Minimum Redesign Phase 2: Early Deployments Interop & Qual Gather multi-vendor qualification data across temperature, BER margin, and liquid cooling manifolds Multi-Vendor Supply Resilience Phase 3: Scale Optical Compute Fabrics Graduate qualified optical engines interconnects at volume. Efficient Cluster Architecture Multi-vendor MSAs eliminate proprietary lock-in, enabling rapid volume scaling and supply chain resilience

p10

• Join the XPO, Open-CPX & OCI MSA • Contribute to the specifications and the roadmap • Pilot XPO/ Open-CPX ready platforms • Bring liquid-cooled, pluggable interconnects into your next 102.4T/204.8T design cycle • Potential OCP discussion areas: − MSA for emerging technologies − Qualification and validation methods − Multi-vendor interoperability Call to Action

p11

Thank You!