DCDeepTech Logo

Rack and Power

OCP · 48V → 800V HVDC · Liquid Cooling

Liquid cooling blueprint for 30–100 kW AI/HPC racks.

This one-page brief summarizes DCDeepTech’s consulting view on rack-level liquid cooling for AI/HPC: how to connect cooling towers, chillers, CDUs and racks; how to choose between DLC, RDHx and immersion; and how to turn OCP-style designs into reliable 30–100 kW liquid-cooled racks.

Rack density
30–100 kW
Power Arch
800V HVDC
Cooling
DLC / RDHx
Target PUE
< 1.15
Typical questions we answer
  • How do I connect my racks to the building chiller loop?
  • When should I use DLC vs RDHx vs immersion?
  • What does a “good” CDU and manifold design look like?
  • How do we monitor, alarm and maintain the liquid loop safely?

Energy Architecture Simulator

Conductor Cross-Section
Copper Busbar
Current (Amps)
2083 A
Critical I²R Heat Loss (Tyranny of Copper)
Rack Power
100,000 W
Line Loss
EXTREME
Efficiency
Low

Advanced Power Solutions: SST + Flywheel

Integrated Solid State Transformer (SST) and Flywheel Energy Storage for unparalleled efficiency and dynamic response in AI/HPC data centers.

SST: High-Efficiency Power Conversion

  • High-frequency power electronics (SiC-based).
  • End-to-end efficiency up to 99\%.
  • Bi-directional energy flow and deep power quality control.

Flywheel: Instant, Safe, Long-Life UPS

  • Microsecond to millisecond response speed.
  • Inherently safe, no thermal runaway risk.
  • Ultra-long life (>20 years) with infinite cycle count.
  • Strong environmental tolerance (unaffected by heat/humidity).

System Advantages for AI Clusters

  • Ultra-fast dynamic regulation for AI/GPU transient load.
  • High power density (800-2800kW impact tolerance).
  • PUE reduction: Overall system efficiency up to 99\%.
  • Green and sustainable: No chemical batteries required.
New Immersive Experience

Global 24/7 Digital Showroom

Step into our virtual consulting space. Explore 3D models of liquid-cooled ORV3 racks, examine CDU piping configurations, and visualize the future of AI infrastructure directly from your browser.

Enter Showroom

Rack & Power Services

Consulting focus on mechanical integrity and energy architecture.

Rack & manifold design

  • Mechanical reinforcement for ORV3/ORW.
  • Reference manifold layouts and hose routing.
  • Material guidance (copper, plating, elastomers).

Export finance advisory – “EFS–LIS–SBLC”

For export projects, our partner AxiomDeta advises on trade finance structure. We recommend the "EFS-LIS-SBLC" model for 90-day payment terms, leveraging Singapore's Enterprise Financing Scheme and Loan Insurance Scheme to issue SBLCs as payment guarantees, minimizing interest costs and risk.

OCP · 48V → 800V HVDC · 液冷

AI / HPC 30–100 kW 機櫃 液冷藍圖。

本單頁簡報總結 DCDeepTech 對 AI/HPC 機櫃級液冷 的諮詢觀點： 如何連接冷卻塔、冰水機、CDU 和機櫃；如何選擇 DLC、RDHx 和浸沒式；以及如何將 OCP 式設計轉化為可靠的 30–100 kW 液冷機櫃。

機櫃密度
30–100 kW
電源架構
800V HVDC
冷卻
DLC / RDHx
目標 PUE
< 1.15
我們常見的諮詢問題
  • 如何將機櫃連接到建築物的冰水機迴路？
  • 何時應使用 DLC、RDHx 或浸沒式冷卻？
  • 「良好」的 CDU 和歧管設計是什麼樣子？
  • 我們如何安全地監控、警報和維護液體迴路？

能源架構模擬器

導體截面積
銅排
電流 (安培)
2083 A
I²R 嚴重熱損 (銅的暴政)
機櫃功率
100,000 W
線路損耗
極高
效率
低

進階電力解決方案：SST + 飛輪

整合固態變壓器 (SST) 和飛輪儲能，為 AI/HPC 數據中心提供無與倫比的效率和動態響應。

SST：高效電力轉換

  • 高頻電力電子技術（基於 SiC）。
  • 端到端效率高達 99\%。
  • 雙向能量流動與深度電能質量調控。

飛輪：即時、安全、超長壽命 UPS

  • 微秒至毫秒級響應速度。
  • 本質安全，無熱失控風險。
  • 超長壽命（>20 年），無限循環次數。
  • 環境耐受性強（不受濕度/溫度影響）。

AI 集群系統優勢

  • 超快速動態調節，應對 AI/GPU 瞬態負載。
  • 高功率密度（800-2800kW 衝擊容忍）。
  • PUE 降低：整體系統效率高達 99\%。
  • 綠色永續：無需化學電池。
New

全球 24/7 數位展廳

走進我們的虛擬諮詢空間。直接從您的瀏覽器探索液冷 ORV3 機櫃、CDU 管線配置的 3D 模型，並預覽 AI 基礎設施的未來。

進入展廳

機櫃與電源服務

諮詢重點：機械結構完整性與能源架構。

機櫃與歧管設計

  • ORV3/ORW 的機械加固。
  • 參考歧管佈局和軟管走線。
  • 材料指導（銅、電鍍、彈性體）。

出口金融顧問 – 「EFS–LIS–SBLC」

針對出口專案，我們的合作夥伴 AxiomDeta 提供貿易金融結構諮詢。我們建議使用「EFS-LIS-SBLC」模式以實現 90 天付款條款，利用新加坡的企業融資計畫和貸款保險，開立 SBLC 作為付款保證，將利息成本和風險降至最低。

Cooling

1. Cooling loop – from cooling tower to rack CDU

A modern liquid-cooled AI rack is part of a three-stage cooling chain: cooling tower and condenser water, chiller and facility water system (FWS), and the technology cooling system (TCS) that serves CDUs and racks.

Data center cooling loop with cooling tower, chiller, CDU and racks

Cooling tower and chiller provide cold water to the Facility Water System (FWS). CDUs couple FWS to the rack coolant loop.

Key design points

  • Clearly separate CWS / FWS / TCS to simplify ownership.
  • Size CDUs for realistic IT load, not just nameplate power.
  • Plan isolation valves and bypass paths for safe maintenance.
  • Instrument temperature, pressure and flow at CDU inlets and outlets for predictive maintenance.
Immersion fluid

For immersion, we work with OCP-inspired dielectric fluids suitable for multi-year operation.

Immersion-specific considerations

  • Evaluate fluid lifetime, compatibility with plastics/seals.
  • Design tanks and service tools for safe lifting of heavy servers.
  • Plan fluid monitoring (moisture, particles, acidity).

2. Liquid cooling options

At 100kW+ per rack, air is not enough. We evaluate three topologies.

Direct-to-Chip (DLC)

Coolant flows through cold plates on CPUs/GPUs. Most precise. Requires robust manifolds.

RDHx

Liquid-cooled radiator on the rear door. Ideal for upgrading existing rooms without touching servers.

Immersion

Submerged servers in dielectric fluid. High efficiency but complex service model.

CDU

The "heart" connecting facility water to rack coolant. Controls temp, pressure and flow.

CDU & Loop Engineering Services

We provide a design-and-blueprint package for rack vendors and operators.

CDU & loop engineering

  • CDU sizing for AI loads & redundancy (N+1).
  • Hydraulic calculations (pressure/head).
  • Alarm logic for leaks and flow loss.

Monitoring & predictive

  • Sensor set placement (T, P, Flow).
  • Dashboards for operators and OEMs.
  • AI-driven anomaly detection models.

1. 冷卻迴路 – 從冷卻塔到機櫃 CDU

現代液冷 AI 機櫃是三階段冷卻鏈的一部分： 冷卻塔和冷凝水、 冰水機與設施水系統 (FWS)， 以及服務於 CDU 和機櫃的 技術冷卻系統 (TCS)。

Data center cooling loop with cooling tower, chiller, CDU and racks

冷卻塔和冰水機提供冷水給設施水系統 (FWS)。CDU 將 FWS 與機櫃冷卻迴路耦合。

關鍵設計要點

  • 清楚區分 CWS / FWS / TCS，以簡化所有權。
  • CDU 尺寸應以實際 IT 負載為準，而非僅標稱功率。
  • 規劃隔離閥和旁路路徑，確保安全維護。
  • 在 CDU 入口和出口測量溫度、壓力與流量，用於預測性維護。
Immersion fluid

對於浸沒式，我們與 OCP 啟發的介電液合作，適合多年運行。

浸沒式專屬考量

  • 評估流體壽命，與塑膠/密封件的相容性。
  • 設計液槽和服務工具，用於安全吊裝重型伺服器。
  • 規劃流體監測（濕度、顆粒、酸度）。

2. 液冷選項

當機櫃功率超過 100kW+ 時，空氣冷卻已不足夠。我們評估三種拓樸結構。

芯片直冷 (DLC)

冷卻液流經 CPU/GPU 上的冷板。最精確。需要強固的歧管。

RDHx (背門式熱交換器)

液冷散熱器安裝在機櫃背門上。是升級現有不碰觸伺服器的機房的理想選擇。

浸沒式

將伺服器浸沒在介電液中。效率極高，但服務模式複雜。

CDU (冷卻分配單元)

連接設施用水到機櫃冷卻液的「心臟」。控制溫度、壓力與流量。

CDU 與冷卻迴路工程服務

我們為機櫃供應商和營運商提供設計和藍圖套件。

CDU 與迴路工程

  • CDU 尺寸設計，考慮 AI 負載與冗餘 (N+1)。
  • 水力計算（壓力/揚程）。
  • 洩漏與流量損失的警報邏輯。

監控與預測性維護

  • 感測器組件佈局（溫度、壓力、流量）。
  • 供營運商和 OEM 使用的儀表板。
  • AI 驅動的異常偵測模型。

Fiber / Connectivity

1. Expanded Beam Optical Interconnect

Optimizing fiber connectivity for hyperscale AI clusters by minimizing maintenance downtime.

Non-Contact Optical Coupling

The innovative ferrule expands the light beam across an air gap, eliminating physical contact between fiber ends. This drastically reduces sensitivity to dust and contamination, a major cause of signal degradation and maintenance in traditional systems.

Accelerating Deployment Speed

The reduced sensitivity to contamination allows for faster, virtually inspection-free deployment. Installation time per connector can be reduced by over 85\% compared to traditional methods, which is critical for scaling large AI/HPC infrastructures.

2. Synergy with Next-Gen DC Architecture

Optical interconnect technology provides key co-benefits when integrated with advanced power (SST/Flywheel) and thermal systems.

High-Temp/Humidity Environments

  • **Contamination Resistance:** Non-contact optics prevent signal degradation from condensation and dust common in hot, humid airflow.
  • **Dual Assurance:** Advanced power systems ensure uninterrupted operation in high heat, while robust fiber maintains signal integrity regardless of ambient humidity, supporting reliable operation in "harsh environments."

Liquid Cooling & Immersion

  • **Blind Mating / Ease of Service:** Non-contact design is ideal for liquid-cooled environments and compact spaces, enabling reliable blind-mating during liquid-loop maintenance.
  • **High Density & Safety:** While advanced power protects chip power supply, superior fiber solutions address high-density cabling and signal routing challenges in liquid-cooled racks.

Advanced Air Cooling (DCHX)

  • **Improved Airflow:** Optical cables are often thinner and support extremely high-density routing, reducing cable pile-up at the rear of the rack. This minimizes airflow resistance and boosts DCHX cooling efficiency.
  • **PUE Contribution:** Reduced fan energy consumption (due to better airflow) and reduced cooling load (due to high temperature tolerance) both directly contribute to achieving PUE targets below 1.2.

Digital Twin & Autonomous Ops

  • **High-Speed Data Pipeline:** AI models demand massive data throughput. Advanced optical interconnects support 400G/800G, ensuring real-time data transmission to AI engines for digital twin analysis.
  • **Data Foundation:** High-speed, high-density fiber provides the data pipe, complementing the rich sensor data streams supplied by advanced power and CDU monitoring systems.

3. Quantum-ready & Software-defined

Quantum-ready rack blueprint

Today's 800V HVDC and liquid cooling work lays the foundation for future quantum systems requiring ultra-stable power and thermal management.

Autonomous monitoring software

Our long-term goal is software-defined operations: AI agents supervising a digital twin of the rack for a self-healing loop (predict -> decide -> execute).

This section focuses on the future integration of high-speed connectivity, optics, and software-defined networks, which are critical for quantum and next-gen AI clusters.

1. 擴束光學互連 (Expanded Beam Optical Interconnect)

優化超大規模 AI 集群的光纖連接，將維護停機時間降至最低。

無物理接觸光學耦合

創新的插芯通過空氣間隙擴大光束，消除了光纖端面之間的物理接觸。這顯著降低了對灰塵和污染的敏感度，這是傳統系統中信號衰減和維護的主要原因。

加速部署速度

降低的污染敏感度使得部署更快，幾乎無需檢查。相比傳統方法，每個連接器的安裝時間可減少超過 85\%，這對於擴展大型 AI/HPC 基礎設施至關重要。

2. 與新一代數據中心架構的協同作用

當光學互連技術與先進電源（SST/飛輪）和散熱系統整合時，可提供關鍵的協同效益。

高溫/高濕環境 (TDC 2.0)

  • **防塵防霉：** 非接觸光學介面在高濕熱氣流中，不會像傳統介面那樣容易因凝露或灰塵導致信號衰減。
  • **雙重保險：** 先進的電源系統保障電力不因高溫中斷，堅固的光纖技術保障信號不因環境中斷。共同支持「惡劣環境運行」的論點。

液冷與浸沒式 (Unibody / Immersion)

  • **盲插/易維護：** 光學互連的非接觸設計非常適合液冷環境（部分型號）或緊湊空間下的盲插操作。
  • **高密度與安全：** 先進的電源技術保護芯片供電，優異的光纖解決了高密度布線的信號引出難題。

新型風冷 (DCHX)

  • **改善風道：** 光學介面支持極高密度的布線，且線纜通常更細，能減少機櫃背後的線纜堆積，減少風阻，提升 DCHX 的冷卻效率。
  • **能效提升：** 光纖減少了風扇功耗（通暢風道），而先進電源技術支持耐高溫運行減少了空調功耗。兩者皆直接貢獻於 PUE < 1.2 的目標。

數位孿生與自主運維

  • **數據高速路：** 光學互連支持 400G/800G 高速互連，確保數據實時傳輸給計算引擎。
  • **數據底座：** 高速、高密度光纖提供數據管道，與先進電源及 CDU 監控系統提供的豐富感測器數據流相互補充。

3. 量子就緒與軟體定義

量子就緒機櫃藍圖

今日的 800V HVDC 和液冷工作為未來需要超穩定電源和熱管理的量子系統奠定了基礎。

自主監控軟體

我們的長期目標是軟體定義運營：AI 代理監控機櫃的數位孿生體，實現自我修復迴路（預測 -> 決策 -> 執行）。

此部分專注於高速連接、光學元件和軟體定義網路的未來整合，這對於量子和下一代 AI 集群至關重要。

Downloads

Rack, Power & Cooling Specifications

OCP PDU Specification

OCP Facebook Open Rack V2 PDU Specification (230V AC version).

Download PDF
CDU Self-Qualification

NVIDIA's CDU Self-Qualification Guideline for cooling systems.

Download PDF
CDU & Loop Checklist

Engineering checklist for safe and efficient CDU selection and loop design.

Download PDF

機櫃、電源與冷卻規格書

OCP PDU 規格書

OCP Facebook Open Rack V2 PDU 規格書 (230V AC 版本)。

下載 PDF
CDU 自我鑑定指南

NVIDIA 針對冷卻系統的 CDU 自我鑑定指南。

下載 PDF
CDU & 迴路清單

用於安全高效的 CDU 選型和迴路設計的工程核對清單。

下載 PDF

Technical FAQ

Structural, Electrical, and Deployment Questions (Q1-Q9)

結構、電氣與部署問題 (Q1-Q9)

Contact DCDeepTech

Ready to discuss your liquid cooling or OCP rack project?

Address

71 Ayer Rajah Crescent, #02-01
Singapore 139951
Email

info@DCdeeptech.com
WhatsApp

+65 8378 5068
Region

Singapore, Taiwan, ASEAN, Brazil.

About DCDeepTech

Founded by senior engineers and financial experts, we serve the OCP ecosystem through DeepTech consulting, supply chain finance solutions, and global marketing. We act as a neutral design office bridging OCP specs and real-world deployment.

Useful Links

AxiomDeta (Engineering Delivery & Manufacturing)

www.axiomdeta.com
HHE (Power Solutions)

www.honghuienergy.cn/en/
3M (Expanded Beam Optical)

3m.com/3M/en_US/p/c/electronics-components...
OCP (Open Compute Project)

www.opencompute.org/