HPC SUMMIT 2027 - Taiwan

Registration & Welcome Coffee 注册及欢迎咖啡

Lobby

Welcome to HPC Summit Taiwan 2027 歡迎來到 HPC Summit Taiwan 2027

Plenary Hall

Kay Wu

Global Business Development Manager

w.media

The AI Infrastructure Stack Is Being Rebuilt: What Comes After the GPU Supercycle? AI 基礎設施架構正在重塑：GPU 超級週期之後，下一波機會在哪裡？

Plenary Hall

AI infrastructure is evolving beyond the GPU. Compute, memory, networking, storage, power and cooling are becoming increasingly interdependent. This keynote examines where the next infrastructure constraints are likely to emerge and how the technology supply chain must evolve to support increasingly powerful AI systems.AI 基礎設施的發展已不再只圍繞 GPU。運算、記憶體、網路、儲存、電力與散熱之間的依存關係日益緊密。本場演講將探討下一個基礎設施瓶頸可能出現在哪裡，以及供應鏈應如何演進，以支援效能持續提升的 AI 系統。

From Chip to Cluster: Where the AI Infrastructure Bottleneck Moves Next 從晶片到叢集：AI 基礎設施的下一個瓶頸將出現在哪裡？

Plenary Hall

Faster accelerators do not automatically create faster AI systems. As compute performance increases, pressure shifts toward memory, networking, storage, power delivery and cooling. This panel explores where the next system-level bottlenecks are emerging and which technologies will be required to overcome them.更快的加速器並不代表整體 AI 系統必然更快。隨著運算效能提升，壓力正逐漸轉向記憶體、網路、儲存、供電與散熱。本場討論將探討下一個系統級瓶頸可能出現在哪些環節，以及哪些技術將成為突破瓶頸的關鍵。

Next-Generation Compute Platforms: Performance Beyond GPU Specifications 下一代運算平台：超越 GPU 規格的整體效能競爭

Plenary Hall

AI infrastructure performance depends on far more than accelerator specifications. This keynote examines how processors, accelerators, memory, interconnects and system architecture must work together to maximise usable compute performance.AI 基礎設施的效能遠不只是由加速器規格決定。本場演講將探討處理器、加速器、記憶體、互連技術及系統架構如何協同運作，才能真正提升可用運算效能。

800G, 1.6T and Beyond: Building the Network for AI Scale 從 800G、1.6T 到下一世代：打造大規模 AI 所需的高速網路

Plenary Hall

Large GPU clusters are driving unprecedented east-west traffic requirements. This keynote examines how 800G, 1.6T, optical technologies and next-generation fabrics are evolving to support larger AI clusters without allowing the network to become the performance bottleneck.大型 GPU 叢集正在產生前所未有的東西向流量需求。本場演講將探討 800G、1.6T、光通訊技術與下一代網路架構，如何支援更大規模的 AI 叢集，並避免網路成為系統效能瓶頸。

Morning Coffee & Exhibition 上午茶歇與展覽交流

Exhibition Floor

The Rack Is Becoming the Product: Designing AI Infrastructure at System Level 機櫃正在成為完整產品：以系統層級重新設計 AI 基礎設施

Plenary Hall

AI infrastructure is shifting from individual servers toward tightly integrated rack-scale architectures. This keynote explores why compute, networking, power distribution and cooling increasingly need to be designed as one integrated system.AI 基礎設施正從單一伺服器走向高度整合的機櫃級架構。本場演講將探討為何運算、網路、配電與散熱必須從獨立元件轉變為整體系統進行協同設計。

Inside the AI Rack: When Compute, Networking, Power and Cooling Must Be Designed Together 深入 AI 機櫃：當運算、網路、電力與散熱必須協同設計

Plenary Hall

High-density AI racks are collapsing the traditional boundaries between IT and facilities. This panel examines how OEMs, ODMs, component suppliers and infrastructure providers must work together earlier in the design process to avoid integration problems later.高密度 AI 機櫃正在打破 IT 與設施之間的傳統界線。本場討論將探討 OEM、ODM、元件供應商與基礎設施供應商如何在設計初期即展開協作，以降低後續系統整合失敗的風險。

Powering 100kW+ Racks: What Changes Inside the Power Chain? 驅動 100kW+ AI 機櫃：整個供電鏈需要如何改變？

Plenary Hall

As AI racks move beyond 100kW, traditional power distribution architectures are being challenged. This keynote examines changes required across power shelves, PSUs, busways, UPS systems and rack-level distribution.當 AI 機櫃功率超越 100kW，傳統配電架構正面臨重大挑戰。本場演講將探討 Power Shelf、電源供應器、母線槽、UPS 系統與機櫃級配電需要如何演進。

Cold Plates, CDUs and Manifolds: Where Liquid Cooling Projects Actually Fail 冷板、CDU 與分流歧管：液冷專案真正容易失敗的環節

Plenary Hall

Liquid cooling projects often fail at the integration layer rather than at the technology itself. This session explores practical challenges involving cold plates, CDUs, manifolds, water quality, controls, commissioning and ownership boundaries.液冷專案真正的問題往往並非冷卻技術本身，而是系統整合。本場演講將探討冷板、CDU、分流歧管、水質、控制系統、測試驗收及責任界面等實際挑戰。

AI Infrastructure Buyers Meet the Supply Chain: What APAC Customers Actually Need 當 AI 基礎設施買家遇上供應鏈：亞太客戶真正需要什麼？

Plenary Hall

Technology roadmaps do not always match deployment realities. This panel brings the buyer perspective into the discussion, focusing on interoperability, deployment speed, support, energy efficiency, supply certainty and total cost of ownership.技術發展藍圖未必與實際部署需求一致。本場討論將從買家的角度出發，聚焦互通性、部署速度、技術支援、能源效率、供應穩定性及總持有成本等實際需求。

Networking Lunch & Exhibition 午餐與展覽交流

Exhibition Floor

Scale-Up vs Scale-Out: The Architecture Battle Behind the AI Factory Scale-Up vs Scale-Out：AI Factory 背後的架構之爭

Plenary Hall

As AI systems grow, infrastructure architects must determine how compute should scale. This keynote examines the trade-offs between scale-up and scale-out approaches and their impact on interconnects, performance, reliability and system cost.隨著 AI 系統規模擴大，基礎設施架構師必須重新思考運算資源應如何擴展。本場演講將比較 Scale-Up 與 Scale-Out 架構在互連、效能、可靠性及系統成本上的不同取捨。

GPU Is Only the Beginning: Memory, Storage and Networking Become the New Constraints GPU 只是起點：記憶體、儲存與網路正成為新的效能瓶頸

Plenary Hall

Adding more accelerators does not guarantee proportional performance gains. This panel explores how memory bandwidth, storage throughput and network congestion increasingly determine whether expensive AI infrastructure can deliver its expected performance.增加更多加速器並不代表效能能夠等比例提升。本場討論將探討記憶體頻寬、儲存吞吐量與網路壅塞，如何決定昂貴的 AI 基礎設施能否真正發揮預期效能。

AI Storage at Scale: Feeding the GPU Without Starving the Cluster 大規模 AI 儲存：如何持續供應 GPU 資料，避免叢集「挨餓」？

Plenary Hall

AI clusters require enormous volumes of data to move continuously across the infrastructure. This keynote examines how flash, distributed storage and data pipelines can be architected to keep GPUs productive and prevent storage from becoming a hidden bottleneck.AI 叢集需要大量資料在基礎設施中持續流動。本場演講將探討快閃儲存、分散式儲存與資料管線應如何設計，以維持 GPU 使用效率，並避免儲存成為隱性瓶頸。

From 100kW to 1MW per Rack: Can Physical Infrastructure Keep Up With Compute? 從每櫃 100kW 邁向 1MW：實體基礎設施能跟上運算密度嗎？

Plenary Hall

AI roadmaps are pushing rack densities far beyond conventional data center design assumptions. This panel examines whether electrical systems, cooling, mechanical infrastructure and facility design can realistically keep pace with next-generation compute.AI 技術發展正將機櫃功率密度推向遠高於傳統資料中心設計假設的水準。本場討論將探討電力系統、散熱、機電基礎設施與設施設計，是否能夠跟上下一代運算需求。

Closing Remarks 閉幕致辭