How JD’s Modular Servers Redefine Data Center Efficiency and Flexibility

In the era of rapid AI and IoT growth, all data processing relies on robust data‑center infrastructure; servers act as the frontline soldiers whose design directly impacts overall performance.

JD.com, leveraging extensive experience in building its own data centers, recently presented a technical open class titled “Modular Design for All‑Scenario Deployment – JD Intelligent Cloud’s Self‑Developed Server Exploration.” Experts Wang Shifeng and Intel cloud architect Zhou Chao discussed the pain points of current IDC servers and introduced JD’s next‑generation modular server concepts, technical breakdowns, and data metrics.

Traditional 2U rack servers suffer from front‑mounted disks and rear I/O, limiting power‑distribution centralization, while many modern servers centralize power but place I/O at the front, causing compatibility issues. These limitations hinder flexible migration and deployment.

The most rigid metric in existing data centers is PUE (Power Usage Effectiveness), often ranging from 1.4 to 3, with older facilities exceeding 3 due to lack of hot‑/cold‑aisle separation. Reducing PUE is a critical goal for future ultra‑large data centers.

Deployment and maintenance are also cumbersome: unpacking, racking, and wiring a server takes about an hour, limiting an engineer to one rack per day. Front‑disk, internal fan, and rear cabling designs require full power‑down for component replacement, preventing intelligent, automated operations.

High‑density layouts still rely on traditional air cooling, which cannot meet the thermal demands of upcoming 350 W CPUs in 1U/2U nodes, prompting a shift toward cold‑plate liquid cooling or immersion cooling.

JD’s experts predict future servers will be highly customizable per client needs and delivered as whole‑rack units to boost delivery efficiency. Data centers will strive for lower PUE, while IDC operations move toward simplification, intelligence, and even unmanned management.

The new modular server introduces a groundbreaking front‑back I/O module that can switch between front and rear I/O, enabling compatibility with standard racks lacking front‑cable management. The design includes a front disk bay, front I/O module, central motherboard, rear PCIe module, and optional rear I/O module, supporting CRPS standard power and a power‑conversion module for centralized rack power management. An integrated water‑cooling module enables hybrid air‑liquid cooling.

Benchmark tests with Seagate showed the JD server achieving near‑100% performance with 14 TB drives (versus 85% for traditional servers) and only a 3% performance loss with 20‑24 TB drives (traditional servers dropped to 7% usable performance).

Key innovations addressing hard‑disk performance degradation include:

Physical separation of fans and disks at opposite ends of the server.

Adoption of a new liquid‑cooling module that lowers fan speed, reducing vibration impact on disks.

Use of Intel’s second‑generation custom CPUs (Icelake), delivering a 30% performance improvement over the first generation.

Overall, JD’s self‑developed server delivers six major optimizations:

New architecture compatible with both rack‑mount and standard servers.

First modular server supporting both front‑ and rear‑I/O.

30% performance boost from Intel custom CPUs.

8% efficiency gain via centralized power system.

50% improvement in cooling efficiency through hybrid cooling design.

5‑10× increase in delivery efficiency.

JD Intelligent Cloud also offers HaaS (Hardware‑as‑a‑Service) integrated solutions combining air‑conditioning, UPS, custom servers, switches, and monitoring software, as well as AI, PASS, and office all‑in‑one machines, extending its hardware expertise to enterprise markets.

Liquid cooling is gaining traction as traditional air cooling struggles with rising component power densities. Intel has heavily invested in liquid‑cooling research, and Intel cloud architect Zhou Chao highlighted JD’s liquid‑cooling design and practice during the technical open class.

The liquid‑cooling system features a primary and secondary loop with plate heat exchangers, a CDU that pumps coolant to server racks for direct CPU cooling, and a secondary loop that returns heat to the CDU for external exchange.

Critical components such as CPU and memory cold plates, tubing, internal liquid containers, and leak detection sensors are designed for optimal performance. A novel dual‑mode node‑level splitter allows seamless switching between air and liquid cooling based on CPU power consumption, using identical form factors for both cooling methods.

To mitigate leak risks, JD employs a liquid‑leak sensing rope placed in coolant channels; voltage changes detected by BMC enable real‑time leak monitoring and remote reporting.