PDF 原檔：報告_OCP2026_AsteraLabs_自帶XPU開放推論機櫃_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_AsteraLabs_自帶XPU開放推論機櫃_20261015_p2.png |
p2 | 真資料圖 | McKinsey 資料中心需求模型長條圖：2025～2030 AI Training 與 AI Inference，2026 兩者約相等，2030 Inference 約 93 對 Training 約 62 |
報告_OCP2026_AsteraLabs_自帶XPU開放推論機櫃_20261015_p8.png |
p8 | 真資料圖 | Orion 機櫃示意：多個 Orion-CT 運算托盤（每個內含 Scorpio X-Series 搭 CPU 與 4 GPU）經中間 Orion-FT 組構托盤互連；右為 CT 與 FT 實體照 |
報告_OCP2026_AsteraLabs_自帶XPU開放推論機櫃_20261015_p9.png |
p9 | 真資料圖 | 多層交換組構拓撲：上下兩排運算托盤、中間一排 Scorpio X-Series 交換器，並有 Orion Rack 實物與托盤排列；標語為擴充至最多 160 GPU |
報告_OCP2026_AsteraLabs_自帶XPU開放推論機櫃_20261015_p13.png |
p13 | 真資料圖 | 左：2x 小交換器（4 GPU P2P 各一）對 1x Scorpio 2 320L（8 GPU P2P）的 Qwen3.5-122B 輸出 token/s 折線（64～512 請求），+50% tokens/s、prefill 重負載 +90%；右：Hypercast 與 INC 示意，集合運算最多加速 2x |
報告_OCP2026_AsteraLabs_自帶XPU開放推論機櫃_20261015_p15.png |
p15 | 真資料圖 | 通用 scale-up 機櫃架構：Orion-CT 用 Scorpio P-Series、Orion-FT 用 Scorpio X-Series，Orion Rack 內可換不同 XPU 加速卡（三色卡示意） |
原始內容
p1
Chris Petersen VP, CTO Office, Astera Labs Choose Your Own XPU: An Open Rack to Maximize Inference Performance and TCO AI CLUSTERS
p2
The Shift to Inference Global AI Data Center Demand by Workload Source: McKinsey Data Center Demand Model: https://www.mckinsey.com/featured-insights/charts/the-future-of-ai-workloads 0 10 20 30 40 50 60 70 80 90 100 2025 2026 2027 2028 2029 2030 AI Training AI Inference Inference spending is projected to surpass training spending for the first time in 2026. Inference infrastructure spending will exceed training by > 50% by 2030.
p3
What Hyperscalers & AI Labs Demand from AI Infrastructure Token Economics Optimize performance per watt and cost per token Scalability Enable diverse accelerators at cluster scale Resilience Maximize uptime and utilization
p4
Universal Scale-up Fabric Solution Universal, Memory Semantic Scale, Bandwidth, Low Latency Maximize GPU Utilization Always-on & Resilient Requirements
p5
Universal Scale-up Fabric Solution Universal, Memory Semantic Scale, Bandwidth, Low Latency Maximize GPU Utilization Always-on & Resilient Requirements PCIe-based, Platform-specific Largest Radix Hypercast & In-Network Compute Multi-plane & Advanced Fabric RAS
p6
AI Fabric Solution Universal Scale-up Fabric Solution Universal, Memory Semantic Scale, Bandwidth, Low Latency Maximize GPU Utilization Always-on & Resilient Requirements PCIe-based, Platform-specific Largest Radix Hypercast & In-Network Compute Multi-plane & Advanced Fabric RAS
p7
Scorpio X-Series 320 Lane Smart Fabric Switch Largest open, memory-semantic switch CPU CPU Orion-CT Compute Tray: PCIe GPU Add-in Cards
p8
Scorpio X-Series 320 Lane Smart Fabric Switch Largest open, memory-semantic switch \ Orion-CT Compute Tray: PCIe GPU Add-in Cards Orion-FT Fabric Tray: High-density scale-up
p9
Scorpio X-Series 320 Lane Smart Fabric Switch Largest open, memory-semantic switch Expand GPU scale-up with multi-level switching fabric – up to 160 GPUs Orion Rack
p10
Mixture-of-Experts (MoE) Makes Larger Models Practical Compute MoE improves efficiency by activating only a small subset of experts for each token instead of the full model. This requires unique multicast groups to target combinations of experts.
p11
Legacy Switches Do Not Support Enough Multicast Groups Compute When only some collectives can be accelerated, the rest fall back to inefficient unicast behavior and become bottlenecks.
p12
Astera Labs’ Hypercast Technology Compute Hypercast is an AI-native multicast technology designed to support both dynamic expert routing and high-frequency collectives.
p13
Size and Intelligence Matter Do more with less: The ultimate differentiator in a constrained environment 0 500 1,000 1,500 2,000 64 128 256 512
Inference Requests Active
1x Scorpio 2 320L Output Tokens Per Second Qwen3.5-122B, Balanced Inference Prefill/Decode 2x Small Switch +50% Tokens Per Second +90% in Prefill heavy workloads ↑ GPU utilization $ Millions rack-scale value Hypercast Data Replication Engine Accelerates one-to-many Compute Accelerate collective operations by up to 2x Faster routing and collectives boost response and Tokens per Second ↑ GPU utilization $ Millions rack-scale value 8 GPU P2P Full Bandwidth GPU GPU GPU GPU GPU GPU GPU GPU CPU CPU 4 GPU P2P 4 GPU P2P GPU GPU GPU GPU CPU CPU GPU GPU GPU GPU Small Switch Small Switch + INC In-Network Compute Accelerates many-to-one
p14
Resiliency at All Levels 99% of your links will work perfectly; the final 1% will take 99% of your deployment time to find and re-seat. Resilient Architecture Lower Opex and maximize uptime Isolated data plane from management/control plane Switch Core µP Signal Faster repair & recovery Identify damaged caps, loose cables, and disconnected pins for every signal cable ` Link Faster product qualification Minimize error rates and quickly isolate link issues Fabric Faster product deployment Mitigate failures with multi-plane routing and advanced fabric RAS GPU GPU GPU GPU GPU GPU At Scale, Rare Events Aren't Rare – James Hamilton
p15
Common Scale-up Rack Architecture Open Protocol High BW & Low Latency Hypercast & INC RAS & Resilience XPU Optionality Leverage common infrastructure across multiple XPUs and generations to reduce TCO Orion-CT Orion-FT Orion Rack Scale-Up GPUs + Memory
p16
Enriching Fabric & Rack-Level Mgmt with COSMOS Managing: System, fabric, and rack-level configuration using open standards (OCP, OpenBMC, DMTF, etc) . Monitoring: Performance metrics, error events, fabric notifications Detect & Resolve Fabric Issues Identify Fabric Bottlenecks Rapidly Configure & Deploy Observability, diagnostics and lifecycle management across the entire fabric
p17
Key Takeaways Astera Labs is delivering AI Your Way with purpose-built rack- scale connectivity. Scorpio and Leo Smart Fabric Solution • Largest open, memory-semantic switch • Full end-to-end rack-scale solution • Collective acceleration via Hypercast and INC • Optimized performance per watt and cost per token • Interop-tested and validated with leading XPUs A full portfolio of copper & optical AI connectivity solutions!
p18
• Now is the time for a common scale-up infrastructure that works with any XPU! • Get involved with OCP Projects • Open Adapter Card (OAC), Hardware Management, RAS, DC-SCM, etc. Learn More: Call to Action Visit Astera Labs in Booth B34 Why Your Mixture-of- Experts Model is Only as Good as Your Fabric AI Your Way. Purpose-built Rack-scale Connectivity. Supercharging Agentic AI: How Leo X-Series Smart Memory Controllers Accelerate Inference with KV Cache Offload
p19
Thank You!