PDF 原檔：報告_OCP2026_Broadcom_主題演講_開放創新擴展AI_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_Broadcom_主題演講_開放創新擴展AI_20261015_p5.png |
p5 | 真資料圖 | Broadcom 乙太網產品線：Tomahawk Ultra 51.2T（HPC/Scale-up）、Tomahawk 6 102.4T（Scale-up/Scale-out）、Jericho 4 51.2T（Scale-across）、Thor Ultra 800G NIC、SUE-T+ESUN |
報告_OCP2026_Broadcom_主題演講_開放創新擴展AI_20261015_p10.png |
p10 | 真資料圖 | Scale-up 架構 SUE-T＋ESUN：ESUN v1.0 規格 2026 年 2 月公布、SUE-T 規格目標 2H2026；ESUN 2025 OCP 宣布，175+ 成員，ESUN 機櫃已量產；成員 logo 含 AMD、Arista、Arm、Broadcom、Cisco、HPE、Marvell、Meta、Microsoft、NVIDIA |
報告_OCP2026_Broadcom_主題演講_開放創新擴展AI_20261015_p12.png |
p12 | 真資料圖 | Ethernet 作為 scale-up fabric：512x100G-PAM4、單跳 scale-up 256、250 ns 延遲、51.2T@64B 77 BPPS，並列 link layer retry、credit-based flow control、in-network collectives |
報告_OCP2026_Broadcom_主題演講_開放創新擴展AI_20261015_p16.png |
p16 | 真資料圖 | 機櫃內銅背板 scale-up 比較：現世代 1.04 Pbps（72 XPU、512x200G PAM4、約 250 kW、Tomahawk6 量產中）vs 可能次世代 2.08 Pbps（576x400G PAM6、約 500 kW），銅距約 2 m、SerDes 7 pJ/bit |
報告_OCP2026_Broadcom_主題演講_開放創新擴展AI_20261015_p20.png |
p20 | 真資料圖 | OCI（Optical Compute Interconnect）：OCI-MSA 2026 年 3 月成立，創始成員 logo 為 Meta、Microsoft、OpenAI、AMD、Broadcom、NVIDIA；目標將 scale-up 從單櫃擴至多櫃與多排 |
報告_OCP2026_Broadcom_主題演講_開放創新擴展AI_20261015_p22.png |
p22 | 真資料圖 | 200G 連結功耗與距離表：DAC 7 pJ/bit、2 m；Integrated OCI 7 pJ/bit、500 m；DR LRO 前板可插拔 23 pJ/bit、500 m |
原始內容
p1
Democratizing AI Infrastructure: Accelerating Innovation at Scale Asad Khamisy Senior Vice President and General Manager, Core Switching Group, Broadcom
p2
Scaling AI Infrastructure Through Open Innovation Asad Khamisy, Ph.D. SVP/GM, Core Switch Group, Broadcom Oct 12th, 2026
p3
100K+ XPU Clusters A Very Very Large Distributed Computing System
p4
Scale-up In Rack Scale-out Across Racks Datacenter1 Datacenter2 Across Racks … Spine Leaf … … … Spine Leaf … … Across Data Centers Scale-out Scale-across AI Scale-up, Scale-out and Scale-across Networking
p5
Jericho 4 51.2 Tbps Tomahawk 6 102.4 Tbps HPC Scale-Up Scale-Out Scale-Across SWITCH Tomahawk Ultra 51.2 Tbps NIC SUE-T + ESUN (OCP) Thor Ultra 800G SAI / SONiC / SDK ESUN / Enhanced RoCE Stack Complete Coverage of HPC and AI with Ethernet
p6
Scale-out Across Racks Datacenter1 Datacenter2 Across Racks … Spine Leaf … … … Spine Leaf … … Across Data Centers Scale-out Scale-across AI Scale-up Scale-up In Rack
p7
Scale-up Connects Tightly-Coupled Compute ` ARCHITECTURE Open Standards PHYSICAL MEDIUM Copper vs Optics
p8
Scale-up Connects Tightly-Coupled Compute ` ARCHITECTURE Open Standards PHYSICAL MEDIUM Copper vs Optics
p9
64 - 144 XPUs Single Rack 1K - 2K XPUs Multiple Racks Today Tomorrow … Common Ethernet Operational Foundation Within a rack → Across racks Electrical + optical connectivity Common operations Telemetry • Diagnostics • Automation Common infrastructure Cabling • Optics • Qualification • Sparing Across the AI fabric Scale-up • Scale-out • Scale-across Ethernet for In-Rack and Multi-Rack Scale-Up Ethernet from Scale-up to Scale-out to Scale-across
p10
• Highly reliable & lossless Ethernet • Efficient support for small messages • Efficient multi-hop scale-up topologies v1.0 spec published February 2026 SUE-T: Co-optimize Transport and Compute • Push vs pull memory access • Memory access ordering model • Transaction packing • Reliability • Load balancing • Scheduling SUE-T spec publication target: 2H2026 XPU A XPU B XPU C OCP SUE Transport Other Accelerator Transport Other Accelerator Transport Ethernet PHY Ethernet Data Link Ethernet Header Ethernet for Scale-up Networking (ESUN) ESUN announced at OCP Global Summit 2025 175+ members joined and growing ESUN racks are now in production Scale-up Architecture: SUE-T + ESUN
p11
Ethernet Scale-up: From Concept to Production! 5+ Hyperscalers running Ethernet Scale-up in production 10+ partners building Scale-up Ethernet based racks Multiple custom & merchant XPU Scale-up Ethernet racks to be in production in the next 6 months
p12
In-Network Collectives Link Layer Retry IP or Optimized Ethernet Headers Credit-Based Flow Control Efficiency Reliability Ultra-Low Latency 250 ns 51.2T@64B 77 BPPS Performance Ethernet Today 512 x 100G-PAM4 Single hop Scale-up: 256 Ethernet as the Scale-up Fabric
p13
Ethernet Inherent Efficiency: Variable Packet Sizes Variable-size packets are inherent to Ethernet • Gives more optimization tools to XPU and system designers • Switch achieves full throughput for all packet sizes Optimized UFH header and Reduced IPG • Next version of OCP ESUN spec Use case Link Efficiency Unpacked 256B posted write 92% 2 packed 256B posted writes 96% 4 packed 256B posted writes 98% Native 2kB block transfer 99%
p14
Scale-up Connects Tightly-Coupled Compute ` ARCHITECTURE Open Standards PHYSICAL MEDIUM Copper vs Optics
p15
In-rack Scale-up ORv3 Rack Copper cable backplane Compute Tray ORv3 Rack Switch Tray Ethernet Switches XPUs NICs Backplane Backplane
p16
In-rack Scale-up: Doubling Bandwidth with Copper Backplane CURRENT GEN 1.04Pbps TOTAL RACK BANDWIDTH 14.4Tbps per XPU 72 XPUs 512x200G PAM4 9-12 Ethernet switches Twinax density: 1024 per OU at 26 AWG XPU to Switch copper reach: ~2m SerDes power: 7 pJ/bit Total rack power: ~250 kW Bandwidth limited by backplane density Production volume with Tomahawk6 today Possible NEXT GEN (2x BANDWIDTH) 2.08Pbps TOTAL RACK BANDWIDTH (2x) 28.8Tbps per XPU 72 XPUs 576x400G PAM6 9 Ethernet switches Twinax density: 1024 per OU at 26 AWG XPU to Switch copper reach: ~2m SerDes power: 7 pJ/bit Total rack power: ~500 kW (2x) 2x Density, 2x Bandwidth
p17
` Scaling to 1K and beyond Expanding Scale-up Interconnect Beyond Single-Rack Limits
p18
Disaggregated Rack Scale-up with Optics XPU Trays Switch Trays ORv3 Rack ORv3 Rack Switch Rack ORv3 Rack ORv3 Rack Switch Rack XPU Rack XPU Rack XPU Rack • Larger scale-up domain (e.g. 576) • Lower power per rack • Reduced weight per rack • Higher cost optics compared to copper • Higher latency across racks (5ns / fiber meter )
p19
Disaggregated Rack Scale-up with Copper and Optics XPU Trays ORv3 Rack Hybrid Rack ORv3 Rack Hybrid Rack ORv3 Rack Hybrid Rack N copper links M fiber links XPU Trays • Preserve large scale-up domain • Breaks scale-up into “low latency domain” and “higher latency domain” • Reduces optics costs ORv3 Rack Hybrid Rack ORv3 Rack Hybrid Rack ORv3 Rack Hybrid Rack
p20
Introducing Optical Compute Interconnect (OCI) Founding Members Industry alignment on 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
p21
Disaggregated Rack Scale-up with OCI XPU Trays Switch Trays ORv3 Rack ORv3 Rack Switch Rack ORv3 Rack ORv3 Rack Switch Rack XPU Rack XPU Rack XPU Rack • Preserve large scale-up domain • Same power as copper • Architecturally defined to enable cost & reliability to approach copper KPI’s
p22
Total Link Power and Reach Comparison - 200G Interconnect Total Link Power (pJ/bit) Reach DAC 7 2m Integrated OCI 7 500m DR LRO Front Panel Pluggable 23 500m Link Interconnect Option XPU Switch SerDes SerDes OCI – the reach of Front Panel Pluggable, the power of Copper
p23
OCP: Open Cluster Designs for AI • Delivers open, multi-vendor blueprints to democratize AI infrastructure. • Modular designs from individual racks (XPU pod) to datacenter fabrics (XOC-N). • V1.0 blueprints published. • Scope includes scale-up, scale-out, front-end, storage, and management networks. • 2nd generation designs under development – please join! 53 Member Companies
p24
Thank you Copyright © 2026 Broadcom. All Rights Reserved. The term “Broadcom” refers to Broadcom Inc. and/or its subsidiaries.
p25
（本頁無可抽取文字，內容在圖中）