Compute Ark: Reconstructing Physical Boundaries in the Blackwell Era and the Generational Evolution of Intelligent Computing Infrastructure
Prologue: When Trillion-Parameter Models Hit the “Thermal Physics Wall”
In the wave of Generative AI (AIGC), the iteration speed of computing power has far exceeded the physical capacity of traditional data centers. With the NVIDIA Blackwell architecture—especially the GB300 NVL72 supernode powered by Blackwell Ultra cores—becoming the industry standard, a cold reality has emerged: single-rack power consumption has officially crossed the physical limit of 120kW-140kW.
This is not just a leap in numbers, but a thermodynamic nightmare. At such extreme power densities, traditional “room-level” air conditioning or simple “all-in-one” liquid-cooled cabinets are no longer sustainable. Severe thermal coupling between racks, the pressure of transformer expansion, and the routing space for thousands of NVLink cables are becoming “physical sieges” hindering intelligent evolution.
Based on years of liquid cooling technology accumulation, Xinke Intelligent officially launches the “Compute Ark” split cluster solution. This is no longer a simple data center renovation, but a paradigm revolution regarding computing power, thermal energy, and space.
Modular Decoupling: The Underlying Logic and Physical Philosophy of Split Architecture
Traditional “integrated” solutions often strive to solve all problems within a single enclosure. However, in medium-to-large deployments (e.g., over 50 supernodes), spatial competition triggers fatal technical compromises. The core of Xinke Intelligent’s solution lies in “decoupling”—using physical isolation to allow power supply, cooling, computation, and operation & maintenance (O&M) to each find their place.
1. IT Computing Container (The Brain): Sovereignty of Computing Density
Consisting of 4 dedicated IT containers, each deploying 12-13 GB300 NVL72 supernodes. Through customized blind-mate liquid-cooled backplanes and high-flow manifolds, a zero-redundancy design within the container is achieved. This maximization of pure computing space enables the construction of a powerful heart with up to 55 EFLOPS (FP4) in a minimal physical footprint.
2. Centralized Power Supply Container (The Heart): “Power Regulator” for AI Loads
The GPU load fluctuations during AI training occur in microseconds. Such transient pulse currents pose a significant impact on traditional grids. Xinke Intelligent’s centralized power supply container integrates HVDC (High Voltage Direct Current) architecture and supercapacitor modules, with a total power supply capacity exceeding 8MW. It is not just a hub for energy distribution, but also a “quality smoother” for electricity, ensuring that the GPU cores always receive rock-solid power support even under extreme load transients.
3. Centralized Cooling Container (The Blood): Precise Allocation of Thermodynamics
Serving as the liquid cooling source station for the cluster, this module features built-in large-scale centralized CDUs (Coolant Distribution Units). Compared to decentralized CDUs, centralized cooling achieves better secondary-side fluid balancing and pressure adaptive regulation. Its precision thermal management system dynamically adjusts pump speed and heat exchange efficiency based on the real-time temperature rise of the GPUs, locking the entire cluster’s PUE at an industry peak below 1.15.
4. Corridor Container (Neural Hub): Redefined Maintenance Space
In a cluster with tens of thousands of GPUs, the scale of NVLink interconnections is comparable to nerve fibers. Two corridor containers serve as the “wiring spine” of the cluster, carrying thousands of DAC copper cables and optical fibers. The recessed integrated corridor not only solves the attenuation problem of high-speed signal transmission between containers but also, through physical isolation, allows O&M personnel to avoid entering the high-temperature, high-noise, and high-voltage IT core areas, achieving a double leap in “data center cleanliness” and “personnel comfort.”
Value Across Cycles: Cooling is Computing Power, Heat is Resource
Under the “Dual Carbon” strategy and strict PUE admission systems, thermal management is no longer an accessory to computing power but its “entry ticket.”
Xinke Intelligent’s solution not only pursues extreme PUE but also explores the business loop of “Computing Power Thermalization.” The cluster possesses up to 90% waste heat recovery capability, providing high-grade hot water at 55°C-70°C. This means that intelligent computing centers are no longer “black holes” swallowing electricity, but “hybrid engines” that can provide green thermal energy for industrial preheating, agricultural greenhouses, and even urban district heating. This path of transforming “waste heat” into “resources” is becoming key for AI infrastructure to meet ESG audits.
Future Outlook: From “Mobile Ark” to the Foundation of “Sovereign AI”
This split design of an “8-container cluster” naturally possesses horizontal scalability. It can be rapidly deployed globally via ocean freight or rail to any corner with power access within weeks.
For enterprises, it is a “strategic supply box” to combat civil construction cycles and quickly seize model opportunities. For nations, it is the optimal physical form for building “plug-and-play, autonomous and controlled” Sovereign AI computing power. Xinke Intelligent is providing global customers with the confidence to face physical limits when dealing with next-generation platforms like Blackwell through this standardized “Compute Ark.”
Conclusion
When physical rules become the shackles of intelligent computing evolution, Xinke Intelligent chooses to break them with technology. From decoupling to reconstruction, from heat dissipation to heat recovery, the “Compute Ark” cluster solution is not only a culmination of liquid cooling technology but also a physical pass to the era of Artificial General Intelligence (AGI).
Xinke Intelligent is building a mobile “Digital Forest” for the next era of computing power.
(Note: Product parameters and solution designs mentioned are based on Xinke Intelligent’s technical standards. Descriptions related to GB300 are based on industry roadmaps; actual performance depends on specific configuration environments.)