報告_OCP2026_Celestica_Broadcom_OpenAI_光運算AI網路_20261015

PDF 原檔：報告_OCP2026_Celestica_Broadcom_OpenAI_光運算AI網路_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

Tareq Bustami, VP Market Technology Manish Mehta, VP of Marketing and Operations Binbin Guan Member of Technical Staff Designing the Next AI Fabric: When compute meets optics 2026 OCP GLOBAL SUMMIT

p2

AI at Work AI adoption is growing alongside gains on selected tasks. ChatGPT weekly users Codex weekly users 15 Sep 2025 27 Feb 2026 700M

900M 1B 31 Aug 2026 27 Feb 2026 8 Apr 2026 1.6M 3M 5M 9 Jul 2026 Astra: selected benchmark scores GPT-5.6 Sol GPT-6 Astra Automation-Bench | workflows 18.1% 41.4% Terminal-Bench 4.0 | coding 37.3% 57.9% Terminal-Bench Science 0.1 22.4% 64.6% Higher is better. Maximum across effort settings; not an equal-cost comparison. OpenAI, 3 Sep 2026. Publication dates; separate user scales. 1 of surveyed workers reported better output speed or quality with AI 75% OCP Global Summit 2026

p3

01 · USER EXPERIENCE Request latency Time to complete an inference request TTLT · time to last token End-to-end latency; TBT is another useful proxy 02 · INFERENCE EFFICIENCY Energy per request Energy needed to process an inference request Energy / token · tokens / joule Also measured as tokens/s/kW Lower latency usually costs more energy per token. Energy is a proxy for cost Compare systems across the full Pareto frontier. Efficiency · Tokens/Joule ↑ Request Latency · TTLT → PARETO FRONTIER Metrics that Matter: Latency and Energy Single user latency energy per token Compare efficiency @ matched user experience OCP Global Summit 2026

p4

Each phase hits a different bottleneck; efficiency only counts if the complete request remains within its end-to-end latency target. 01 Prefill Encode context FLOPs + attention Attention-heavy and primarily compute- bound. Low memory-BW demand; communication is easier to schedule smoothly. Compute HIGH Memory BW LOW Comms SMOOTH 02 Draft model Speculate Network latency Small models at ultra-low batch for latency serving—and low batch elsewhere. Low network bandwidth, but extreme latency sensitivity. Model SMALL Batch ULTRA-LOW Comms LATENCY BOUND 03 Spec-verify Decode Attention + HBM BW Attention less memory bound with GQA and spec tokens. At low latency, MoE is memory-bandwidth hungry and communication arrives in bursts. Attention COMPUTE MoE HBM BW Comms BURSTY A Request Spans Three Distinct Hardware Regimes What matters is requests/second/watt at the required SLA latency OCP Global Summit 2026

p5

HBM slice 1 local view Core slice 1 HBM slice 2 local view Core slice 2 HBM slice 3 local view Core slice 3 HBM slice 64 local view Core slice 64 … General NoC · Lower bandwidth, higher latency · More flexible Scale-up Ethernet bridge FASTEST LOCAL PATH Pair each core slice with an HBM slice to expose a low-latency, high-bandwidth local view. PRESERVE FLEXIBILITY Shared NoC for general comms and access to scale-up network. Collective network · High bandwidth, low latency SPECIALIZED CROSS-CORE Collective network optimizes bandwidth and latency of common comms patterns. Co-design the Architecture to Address the Bottleneck Explicit placement and distributed control prevent data movement/synchronization from dominating execution Keep common operands local; use the shared path only when necessary OCP Global Summit 2026

p6

OAI Jalapeño: from Chip to Scale-up System Host + accelerator rack pair • Higher bandwidth for tensor parallelism • Lower bandwidth for expert parallelism • Low latency for both OpenAI, Hot Chips 2026 “Half flattened” two-level Clos topology OCP Global Summit 2026

p7

Scale-up In Rack Scale-out Across Racks Data center 1 Data center 2 Across Racks … Spine Leaf … … … Spine Leaf … … Across Data Centers Scale-out Scale-across AI Scale-up, Scale-out and Scale-across Networking OCP Global Summit 2026

p8

Critical Optical Interconnect Building Blocks Multi-mode, short reach Single-mode, long reach VCSEL EML Single-mode for Silicon Photonics and CPO CWL Architecture, Reliability, and Power leadership, AI Scale-out CPO Architecture, Reliability, Latency and Power leadership, AI Scale-up OCI XPU Switch DSP Optics OCP Global Summit 2026

p9

Progression of Highly Integrated Photonics Technologies Next Gen 2028 OCI 102.4T 2026 TH6 Davisson 200G/lane 51.2T 2024 TH5 Bailly 100G/lane 25.6T 2022 TH4 Humboldt 100G/lane OCP Global Summit 2026

p10

Scaling Highly Integrated Photonics: Challenges to Overcome Auto dispenser with loader and unloader Detachable fiber connector TH5-Bailly FOWLP structure Advanced Packaging • High-density and high-quality PIC to EIC attach required • Must optimize signal integrity • Enablement for 400G direct drive Optical Connectivity • More than 1k fibers connecting to each chip package • Detachable fiber connectivity to enable package reflow Volume Manufacturing • Must deliver industry leading yield and reliability through automation • Enable system partners with automated assembly and test OCP Global Summit 2026

p11

Progress Towards Scale and Automation Support Frame In-line OGP Delatch Z-Springs Bucket Demonstration of purpose-built equipment OCP Global Summit 2026

p12

Introducing Optical Compute Interconnect (OCI) Founding Members Industry alignment around an optical scale-up. Future AI infrastructure built on open standards. Architecture that delivers the efficiency required to scale AI infrastructure. Optical Scale-Up Est. March 2026 Move Scale-up from In-Rack to Multi-Rack & Multi-Row OCP Global Summit 2026

p13

Does the OCI-MSA Address Scaling? • AI pace of innovation likely faster than “standard” Ethernet AUI interfaces • Future larger AI cluster scale-up domains will require highest radix switches Bandwidth Spec Timeline Electrical Optical Linear Comment 200Gbps Today 100Gbaud PAM4 100Gbaud PAM4 YES 400Gbps 2027? PAM4/6/8, SBD? 200Gbaud PAM4 Maybe? IC area & pJ/b diminishing returns Maybe cabled backplane support 800Gbps

2030? ? ? Unlikely OCI GEN1 Today NA, wide parallel 4λ x 50Gbps NRZ, DWDM, bidi NA GEN2 spec @ 400Gbps = 2027 GEN3 spec @ 800Gbps = 202x OCP Global Summit 2026

p14

Designed, integrated and tested as one system • Top-of-rack switch • AI fabric switch blades • AI compute servers • NVMe storage • Cooling distribution unit • Power and battery-backup shelves 800G and 1.6T AI fabrics, 112G/224G signaling, qualified optics and connectivity Advanced liquid cooling, cable backplanes, and fully integrated modular AI racks, including OCP ORv3 Dense enterprise-class expansion with high- availability PCIe Gen 5 controllers SONiC-based software, firmware, and lifecycle services The Complete AI Rack System NETWORKING AI COMPUTE STORAGE SOFTWARE & RACK INTEGRATION Four disciplines, engineered in unison OCP Global Summit 2026

p15

Annual Qualification windows are shrinking — no room for iteration loops Months to Qualify 36 24 18 36 months · 100G / 400G 24 months · 800G 18 months · 1.6T 9 The Squeeze: Tech Acceleration Coupled with Faster Qualification ~5,000 Passive copper cables in one scale-up rack ~1 MW Liquid-cooled racks reshape thermal, mechanical and cable design 224G SI challenges at high data rates 1.6T Copper reach shrinks; transition to optical ~4,000 lb Rack weight today; 8,000 lb design target ahead Solution qualification is the new critical path COMPOUNDING COMPLEXITY New AI accelerators from major silicon vendors OCP Global Summit 2026

p16

2027 – 2030 Migration • Transition to all optical • CPO-based network switches & XPU • OCI slow & wide optical interfaces in scale-up • BiDi multi-wavelength channels over fiber • Compute rack - optics on a separate optical repeater tray or directly on compute trays Transition from Copper to Optical Interconnects Enables Large Scale-up Domains SCALE-OUT SCALE-UP CPO Scale-up Switches OCI slow & wide CPO Scale- out Switches OCP Global Summit 2026

p17

At Scale Manufacturing: CPO Development 17 Category Key Activity Design Schedule Project Kickoff Mockup Build EVT Prototype Build Manufacturing Development FAU / CPO Dev’t Samples SMT Preparation SMT Optimization DOEs Standard Connector Inspection & Cleaning FAU / CPO Inspection and Cleaning FAU Assembly / Loopback Testing SMT Optical Validation Line Layout Test Development Hot Gas Rework Manufacturing Process Readiness - - Front End - - Assembly - - Test EVT Build OCP Global Summit 2026

p18

Introducing Celestica’s DS6005 Open CPO Platform Solution • Celestica plans to contribute the design to the OCP Community • For more information: www.celestica.com/DS6005 • Visit Celestica booth C2 to see the DS6005 AI Fabric solution based on Broadcom’s Tomahawk 6 – Davisson CPO chipset OCP Global Summit 2026

p19

OCP Community: Let’s partner to design the next AI Fabric