HPC SUMMIT 2027 - Singapore

Registration & Welcome Coffee 注册及欢迎咖啡

Lobby

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

Plenary Hall

James Loggie

Program Manager

w.media

APAC’s AI Infrastructure Buildout: From Buying GPUs to Building AI Factories 亞太 AI 基礎設施建設：從採購 GPU 到打造 AI Factory

Plenary Hall

AI infrastructure investment is moving beyond accelerator procurement. Organisations must now integrate compute, networking, storage, power, cooling and software into production environments capable of supporting real workloads at scale. This opening keynote examines how APAC is moving from AI experimentation toward the construction of full-scale AI factories.AI 基礎設施投資已不再只是採購加速器。企業現在必須將運算、網路、儲存、電力、散熱及軟體整合成能夠大規模支援實際工作負載的生產環境。本場開幕演講將探討亞太地區如何從 AI 實驗階段邁向完整 AI Factory 的建設與部署。

The AI Factory Is Not a Data Center: Rethinking Infrastructure for Accelerated ComputeAI Factory 不只是資料中心：為加速運算重新思考基礎設施

Plenary Hall

Traditional data centers were designed around relatively predictable IT loads. AI clusters introduce extreme power density, east-west network traffic, liquid cooling and highly variable workloads. This panel examines how data center operators, system integrators and technology providers must redesign infrastructure around accelerated computing rather than simply fitting GPUs into traditional facilities.傳統資料中心是以相對可預測的 IT 負載為基礎設計，而 AI 叢集則帶來極高功率密度、大量東西向網路流量、液冷需求及高度變動的工作負載。本場討論將探討資料中心營運商、系統整合商與技術供應商如何以加速運算為核心重新設計基礎設施，而非只是將 GPU 放入傳統資料中心。

Designing the AI Cluster: Integration Starts Before the GPU Arrives 設計 AI 叢集：系統整合必須在 GPU 到場之前開始

Plenary Hall

Many deployment problems begin long before hardware installation. Network topology, storage architecture, power availability, cooling design and orchestration must be planned as one system. This keynote examines the critical design decisions that determine whether an AI cluster performs as expected once commissioned.許多 AI 基礎設施問題早在硬體安裝之前就已形成。網路拓撲、儲存架構、電力供應、散熱設計與工作負載調度必須作為同一個系統進行規劃。本場演講將探討哪些關鍵設計決策會決定 AI 叢集在正式啟用後是否能達到預期效能。

The AI Network Fabric: Connecting Thousands of Accelerators Without Losing PerformanceAI 網路 Fabric：如何連接數千個加速器而不犧牲效能？

Plenary Hall

As clusters scale, networking becomes one of the largest determinants of usable compute performance. This keynote explores fabric architecture, bandwidth, congestion management, latency and optical connectivity required to maintain performance across increasingly distributed AI systems.隨著叢集規模持續擴大，網路正成為決定實際可用算力的關鍵因素之一。本場演講將探討網路 Fabric 架構、頻寬、壅塞管理、延遲與光纖互連技術，如何確保大型分散式 AI 系統維持穩定效能。

Morning Coffee & Exhibition 上午茶歇與展覽交流

Exhibition Floor

100kW Is Only the Beginning: Rebuilding the Data Center Around AI100kW 只是開始：圍繞 AI 重新設計資料中心

Plenary Hall

AI infrastructure is pushing rack densities from tens of kilowatts toward hundreds of kilowatts and potentially beyond. This session examines the downstream implications for electrical architecture, transformers, switchgear, UPS systems, cooling plants and data center capacity planning.AI 基礎設施正在將單櫃功率密度從數十 kW 推向數百 kW，未來甚至可能更高。本場演講將探討這一變化對電氣架構、變壓器、開關設備、UPS、冷卻系統及資料中心容量規劃所帶來的連鎖影響。

Compute Meets Facilities: Who Owns the AI Infrastructure Architecture? 當運算遇上設施：誰應該主導 AI 基礎設施架構？

Plenary Hall

The traditional boundary between IT and facilities is disappearing. GPU selection can affect power architecture. Cooling design can influence server configuration. Network topology can alter physical layouts. This panel examines how system integrators, IT architects, MEP consultants, data center operators and technology vendors should divide responsibilities when every infrastructure layer is interconnected.IT 與設施之間的傳統界線正在消失。GPU 選擇會影響供電架構，散熱設計會影響伺服器配置，而網路拓撲亦可能改變機房實體布局。本場討論將探討當每一層基礎設施高度互聯時，系統整合商、IT 架構師、機電顧問、資料中心營運商及技術供應商應如何重新分工。

Brownfield AI: Turning Existing Data Centers into AI-Ready Infrastructure Brownfield AI：將既有資料中心升級為 AI-Ready 基礎設施

Plenary Hall

APAC cannot build entirely new data centers for every AI deployment. Existing facilities will need to accommodate higher rack densities, liquid cooling and substantially greater network capacity. This keynote examines the technical and operational challenges involved in converting conventional facilities into AI-ready environments.亞太地區不可能為每一個 AI 部署專案重新興建資料中心。大量既有設施必須升級以支援更高機櫃功率密度、液冷及更大的網路容量。本場演講將探討將傳統資料中心改造成 AI-ready 環境時所面臨的技術與營運挑戰。

Liquid Cooling Integration: The Interface Between the Server and the Building 液冷系統整合：伺服器與建築基礎設施之間的新介面

Plenary Hall

Liquid cooling connects IT equipment directly with mechanical infrastructure, creating new engineering and operational dependencies. This session explores CDU placement, piping, water management, redundancy, monitoring, maintenance and commissioning from a complete system integration perspective.液冷技術將 IT 設備直接連接至機械基礎設施，形成全新的工程與營運依存關係。本場演講將從完整系統整合角度探討 CDU 配置、管路、水管理、備援、監控、維護及測試驗收等關鍵問題。

The Integration Gap: Why AI Infrastructure Projects Fail Between VendorsAI 基礎設施的整合缺口：為什麼問題往往發生在供應商之間？

Plenary Hall

Individual technologies can perform exactly as designed while the complete system still fails to meet expectations. This panel examines the integration gaps between compute, network, storage, power, cooling and software vendors, and how project owners can establish clearer architecture ownership, testing standards and accountability.即使每項單一技術都按照設計正常運作，整體 AI 系統仍可能無法達到預期效能。本場討論將探討運算、網路、儲存、電力、散熱及軟體供應商之間常見的整合缺口，以及專案業主應如何建立更明確的架構責任、測試標準與問責機制。

NETWORKING Networking Lunch & Exhibition 午餐與展覽交流

Exhibition Floor

GPU Utilisation Is the New PUE: Measuring the Economics of AI Infrastructure GPU 使用率就是新的 PUE：重新衡量 AI 基礎設施經濟效益

Plenary Hall

A low PUE does not create value if expensive accelerators remain underutilised. As AI infrastructure costs rise, organisations must measure how effectively compute resources are converted into productive workloads. This keynote explores GPU utilisation, scheduling, job completion and infrastructure productivity as emerging measures of AI infrastructure efficiency. 如果昂貴的加速器長時間處於閒置狀態，即使 PUE 很低，也未必能創造真正價值。隨著 AI 基礎設施成本持續上升，企業必須衡量運算資源究竟有多少真正轉化為有效工作負載。本場演講將探討 GPU 使用率、排程、工作完成效率及基礎設施生產力等新一代 AI 效率指標。

GPU-Rich but Compute-Poor: Why Expensive AI Infrastructure Still Underperforms GPU 很多，真正算力卻不足：為什麼昂貴 AI 基礎設施仍然效能低落？

Plenary Hall

Buying more GPUs does not guarantee more usable compute. Software inefficiencies, poorly designed fabrics, storage bottlenecks, scheduling conflicts and insufficient data pipelines can leave accelerators idle. This panel explores why utilisation often falls below expectations and how infrastructure teams can improve end-to-end productivity.購買更多 GPU 並不代表能獲得更多真正可用的算力。軟體效率、網路架構、儲存瓶頸、排程衝突及資料管線不足，都可能導致昂貴的加速器長時間閒置。本場討論將探討 GPU 使用率為何經常低於預期，以及基礎設施團隊如何提升端到端運算效率。

One AI Stack, Many Bottlenecks: Finding What Is Really Slowing Your Cluster 一套 AI 架構，多個潛在瓶頸：真正拖慢叢集的是什麼？

Plenary Hall

When AI applications underperform, the GPU is often blamed first. The actual constraint may be memory, storage, networking, orchestration, cooling or even power throttling. This session presents a system-level approach to diagnosing performance problems across the entire AI infrastructure stack.當 AI 應用效能不佳時，GPU 往往最先被懷疑。然而真正的限制可能來自記憶體、儲存、網路、工作負載調度、散熱，甚至電力限制。本場演講將從完整系統層級探討如何診斷 AI 基礎設施中的實際效能瓶頸。

Training Here, Inference There: Building APAC’s Distributed AI Infrastructure 訓練在這裡、推論在那裡：打造亞太分散式 AI 基礎設施

Plenary Hall

The optimal location for AI training may not be the optimal location for inference. Training requires massive compute and power, while inference may prioritise latency, connectivity, data sovereignty and proximity to users. This closing panel examines how APAC could develop a distributed AI infrastructure architecture across Singapore and other regional markets.最適合 AI 訓練的地點未必最適合 AI 推論。訓練通常需要大量運算與電力，而推論可能更重視延遲、連接能力、資料主權及與使用者的距離。本場閉幕討論將探討新加坡及其他亞太市場之間，如何形成分散式 AI 基礎設施架構。

Closing Remarks: Singapore as APAC’s AI Infrastructure Integration Hub 閉幕致辭：新加坡作為亞太 AI 基礎設施整合樞紐

Plenary Hall