KT Cloud is collaborating with specialized companies in power and cooling to enhance the competitiveness of its Artificial Intelligence Data Center (AIDC). As the introduction of high-performance GPUs leads to a surge in power consumption and heat generation in data centers, KT Cloud aims to combine its operational capabilities with the expertise of these companies to establish a stable, high-density AI infrastructure and improve supply speed.
On November 1, KT Cloud unveiled its AI Innovation Center at the Mokdong 2 Data Center in Yangcheon-gu, Seoul, and introduced its AIDC business strategy and the status of key technology verifications. KT Cloud is working with LS Electric and Vertiv from the equipment design phase and plans to apply the verified technologies to the construction of new data centers.
The collaboration with specialized companies is crucial as AIDC requires infrastructure that differs from simple expansions of existing data centers. GPU servers used for AI training and inference consume more power and generate more heat than traditional CPU-based servers. Consequently, the design of the entire data center must be altered, including power supply facilities, cooling systems, floor load in server rooms, and networks.
Bang Jae-won, head of the AIDC business development team at KT Cloud, stated, "The power consumption per rack in existing data centers was around 2.2 kW in the past, but it has recently increased to between 30 and 150 kW, with projections suggesting it could reach 600 kW in the future. As power density increases, so does the heat generated, making it difficult to rely solely on traditional air cooling methods, thus highlighting the importance of liquid cooling technology."
The AI Innovation Center showcased equipment designed to respond to these technological changes. A notable technology is Direct-to-Chip (D2C) liquid cooling, which involves closely attaching a cold plate to the GPU chip surface and circulating coolant inside to remove heat. This method manages the heat of high-heat servers more efficiently than traditional air cooling.
Another interesting development is a device developed by KT Cloud that simulates the heat generated by high-performance GPU servers without actually installing them. This allows for the verification of cooling equipment performance without the burden of repeatedly testing expensive hardware. A KT Cloud representative explained, "This device is not just about generating and cooling heat; it continuously replicates the conditions of actual servers, including the pressure, flow, and temperature of the coolant."
The center also introduced immersion cooling technology, which removes heat by submerging servers in an insulating cooling liquid, as well as rack power infrastructure for high-density AI servers and RoCEv2-based AI networks. Additionally, it is testing automation technologies for data center operations, such as the AI-based power control solution 'Pathfinder' and 'DIMS Insight,' which detects anomalies and predicts failures. The use of autonomous inspection robots to collect temperature and environmental data in server and power rooms is also being pursued.
KT Cloud has leveraged over 25 years of experience in building and operating data centers as the foundation for its AIDC business. However, it recognizes that constructing data centers typically takes 3 to 4 years, from securing land to power supply, permits, design, construction, and procurement of major equipment, which limits its ability to respond to rapidly increasing market demand solely with its own capabilities.
To address this, KT Cloud is collaborating with LS Electric on power infrastructure and direct current power technologies necessary for modular data centers. With Vertiv, it is pursuing design optimization that comprehensively considers power and liquid cooling facilities. The plan is to apply a modular approach, where power facilities are pre-assembled and manufactured off-site before installation, thereby reducing construction time. This method is set to be implemented at the Gunsan AIDC.
KT Cloud also aims to expand its supply of new data centers. According to its business plan announced in September, it targets to secure over 1 GW of additional AIDC infrastructure by 2031, separate from existing data centers, with new projects underway in Cheongna and Ansan.
Beyond supplying data center buildings and power and cooling facilities, KT Cloud plans to enhance its integrated services, ensuring the stable operation of customers' GPU servers and connectivity between different data centers and clouds. It is also considering a GPU service model for customers who find it challenging to purchase expensive GPUs directly.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.