AI 伺服器電源架構概覽

從電網到晶片：全方位電力生態系統

AI 運算強度的前所未有激增，正從根本上重塑資料中心的電力架構，加速了從標準 3 kW 伺服器向 100 kW+ 高密度電源機架 (Power Shelf) 及 MW 級高壓直流 (HVDC) 機櫃的轉型。隨著業界領先的**「電網至核心 (Grid-to-Core)」**策略演進，驗證這一完整的電力鏈——目前正朝向 800V 直流架構邁進，以支援需電量高達 10,000 安培的處理器——需要更高階的高精密測試儀器。

在這種高功率環境下，傳統的耗散型負載 (Dissipative Loads) 已難以負荷；它們會產生大量的廢熱與噪音，且無法提供現代 AI 工作負載所需的微秒級 (Microsecond-level) 動態響應能力。REGATRON 提供了關鍵解決方案：一套全功率回饋 (Fully Regenerative)、高動態且具備矩陣式配置 (Matrix-configurable) 的電力測試生態系統，專為滿足現代 AI 工廠對效率與擴充性的嚴苛要求而設計。

From Grid to Chip: A Holistic Power Ecosystem (H1)

The unprecedented surge in AI computational intensity is fundamentally reshaping data centre power architectures, accelerating the transition from standard 3 kW servers towards high-density 100 kW+ power shelves and megawatt-scale HVDC racks. As industry-leading "Grid-to-Core" strategies evolve, validating this entire power chain—now shifting towards 800V DC architectures to support processors requiring up to 10,000 Amperes—demands a sophisticated class of high-precision test instrumentation. Traditional dissipative loads have become untenable for these high-power environments; they generate excessive waste heat and noise while failing to provide the microsecond-level dynamic performance necessitated by modern AI workloads. REGATRON provides the essential solution: a fully regenerative, high-dynamic, and matrix-configurable power testing ecosystem designed to match the efficiency and scalability requirements of the modern AI factory.

矩陣式擴充性與前瞻性設計

隨著 AI 電力需求持續攀升，REGATRON 提供的模組化測試架構，旨在隨技術進步同步成長。其 G5 (直流) 與 TC.ACS (交流) 系統皆支援單機無縫整合，可組成並聯、串聯或混合模式的矩陣，總容量最高可達 5,000 kW 以上，電流則超過 20,000 A。這種設計讓企業能從單一 54 kW 模組開始進行初步元件驗證，隨後再根據需求擴充至數 MW 級系統，以進行完整的「從電網到晶片」機櫃級驗證。

卓越的軟體定義控制技術

相較於傳統的電阻式負載，G5 與 TC.ACS 系統是高度先進的軟體定義儀器 (Software-Defined Instruments)。其內部控制迴路 (Control Loops) 具有極佳的適應性，即使面對複雜的線路阻抗 (Cabling Impedances) 或具挑戰性的待測物 (DUT) 特性，仍能確保高度穩定性。這確保了測試數據能真實反映電源供應器 (PSU) 或電源機架 (Power Shelf) 的實際效能，完全不受測試設備本身局限性的影響。

最佳化營運效率與總體持有成本 (TCO)

對於 24/7 全天候運行的測試任務而言，能耗是最大的營運支出。REGATRON 的功率回饋 (Regenerative) 技術確保測試過程中所使用的能量（涵蓋交流與直流階段），高達 95% 可被回收並回饋至廠區電網。全方位的總體持有成本 (TCO) 分析證實，與高電費、高散熱成本的傳統氣冷耗散型負載 (Dissipative Loads) 相比，這種效率能提供顯著更快的投資報酬率 (ROI)，並降低環境足跡。

Matrix Scalability and Future-Proofing

As AI power requirements continue to escalate, REGATRON provides a modular testing architecture designed to grow alongside technological advancements. Both the G5 (DC) and TC.ACS (AC) systems allow for the seamless integration of units into parallel, series, or mixed-mode matrices, reaching capacities of 5,000+ kW and currents exceeding 20,000 A. This approach enables an organisation to commence with a single 54 kW module for initial component validation and subsequently expand to a multi-megawatt system for full "Grid-to-Chip" rack verification.

Sophisticated Software-Defined Control

In contrast to conventional resistive loads, the G5 and TC.ACS systems are highly advanced software-defined instruments. The internal control loops are remarkably adaptable, ensuring high stability even when faced with complex cabling impedances or challenging Device Under Test (DUT) characteristics. This ensures that the resulting test data reflects the authentic performance of the PSU or Power Shelf, entirely independent of the limitations of the test equipment.

Optimal Operational Efficiency and TCO

For 24/7 testing operations, energy consumption represents the most significant operational cost. REGATRON’s regenerative technology ensures that up to 95% of the energy utilised during testing—across both AC and DC stages—is recovered and returned to the facility’s grid. A comprehensive Total Cost of Ownership (TCO) analysis confirms that this efficiency provides a notably faster Return on Investment (ROI) and a reduced environmental footprint compared to the high electricity and cooling costs of traditional air-cooled dissipative loads.

「從電網到晶片」電力基礎設施概覽 (H2)

為了支援次世代 AI 工廠，電力基礎設施正經歷三個不同的演進階段，以有效應對持續攀升的 GPU/TPU 能耗需求：

第一代：整合式機櫃 (100–200 kW)

運算模組與電源模組共用單一機殼。由多組電源供應器 (PSU) 透過共同的匯流排 (Busbar) 提供 50V 直流電，支援最高約 200 kW 的機櫃功率。雖然這是目前的標準設計，但隨著運算密度增加，這種整合式架構終將超出空間與散熱的極限。

第二代 A：電力側櫃 (Power Sidecar) 與高壓直流 (HVDC) (200 kW – 1 MW)

為克服空間限制與配電損耗，電源組件（包括 PSU、電池備援單元 BBU 與電容備援單元 CBU）被移至專用的電力側櫃 (Power Sidecar)。該系統將 400V 或 800V 高壓直流電 (HVDC) 配送至 IT 機櫃，再經由側邊安裝的轉換級 (Step-down stage) 降壓至 50V。此架構能最大化運算容量，同時減少能量浪費。

第二代 B：MW 級固態變壓器 (SST) 與直接轉換 (MW+)

為了達到兆瓦 (MW) 級的極致效率，透過固態變壓器 (SST) 將 34.5 kV 的市電直接轉換為 800V HVDC 微電網，省去了多餘的轉換層級。最終降壓程序透過中間匯流排轉換器 (IBC) 直接整合於伺服器托盤 (Server Tray) 內，將從電網到處理器核心的損耗降至最低。

Overview of Grid-to-Chip Power Infrastructure (H2)

To support next-generation AI factories, power infrastructure is evolving through three distinct stages to manage escalating GPU/TPU energy demands efficiently:

Generation 1: Integrated Rack (100–200 kW)

Compute and power modules share a single enclosure. Multiple PSUs deliver 50V DC via a common busbar, supporting rack power up to ~200 kW. While currently standard, increasing densities eventually exceed the spatial and thermal limits of this integrated design.

Generation 2A: Power Sidecar & HVDC (200 kW – 1 MW)

To overcome space constraints and distribution losses, power components (PSUs, BBUs, and CBUs) are moved to a dedicated 'Power Sidecar'. This distributes 400V or 800V High-Voltage DC (HVDC) to the IT rack, where a final side-mounted conversion stage steps it down to 50V. This architecture maximises compute capacity while reducing energy waste.

Generation 2B: MW-Scale SST & Direct Conversion (MW+)

For megawatt-scale efficiency, Solid-State Transformers (SST) convert 34.5 kV utility power directly into an 800V HVDC microgrid, eliminating redundant conversion stages. Final voltage step-down is integrated within the server tray via Intermediate Bus Converters (IBC), minimising losses from the grid to the processor core.

PSU: Power Supply Unit	VRM: Voltage Regulator Module
BBU: Battery Backup Unit	SSCB: Solid State Circuit Breaker
CBU: Capacitor Backup Unit	SST: Solid State Transformer
IBC: Intermediate Bus Converter