A Total of 7,656 NVIDIA GPU B200s Provided

"High Performance Maintained with Liquid Cooling System"

"Advanced Power, Cooling, Network, and Server Design"

"NHN FactoryX Seoul, which provides a total of 7,656 NVIDIA Graphics Processing Unit (GPU) B200 resources, is the largest facility in Korea and employs a water-cooling system to reliably manage heat generation from high-performance GPUs."


The interior view of "NHN FactoryX Seoul," which provides a total of 7,656 NVIDIA Graphics Processing Unit (GPU) B200 resources. Provided by NHN Cloud.

The interior view of "NHN FactoryX Seoul," which provides a total of 7,656 NVIDIA Graphics Processing Unit (GPU) B200 resources. Provided by NHN Cloud.

View original image
[Report] Inside NHN Factory X Seoul, the Largest Hub of Over 7,000 High-Performance GPUs View original image

With a peak daytime temperature of 38 degrees Celsius during the heatwave on August 7, inside NHN Cloud's artificial intelligence (AI) data center, NHN FactoryX Seoul, located beside an apartment complex in Yangpyeong-dong 3-ga, Seoul, high-performance GPUs were humming busily in a cool environment. The 10-story data center that houses NHN FactoryX Seoul is comprised of data halls from the 3rd to the 10th floor, with NHN Cloud currently utilizing floors 3 through 6. Floors 3 to 5 are allocated for the operation of government AI computing resources, and on the 3rd and 4th floors, 4,080 GPUs have been linked together into a single integrated cluster environment. This allows the large-scale GPU resources to be used as if they are one computational environment.


Upon entering the 6th floor of the data center, the striking 8-meter-high ceilings stand out. The higher the ceiling, the easier it is to achieve cooling, so the height was intentionally designed to aid in dissipating the thermal load generated by GPU servers and storage equipment alike.


Last year, NHN Cloud was selected as a key participant in the Ministry of Science and ICT's "GPU Procurement, Construction, and Operation Support Program." Leveraging the 7,656 NVIDIA GPU B200s secured through this project, the company established the NIPA-CL1 cluster with 4,080 GPUs (510 nodes) and the NIPA-CL2 cluster with 2,040 GPUs (255 nodes). Notably, NIPA-CL1 ranked 20th in the world and the highest in Korea in the June "Global Supercomputer TOP500" ranking.

View of the upper water-cooling pipes at NHN FactoryX Seoul. Provided by NHN Cloud.

View of the upper water-cooling pipes at NHN FactoryX Seoul. Provided by NHN Cloud.

View original image

NHN FactoryX Seoul is distinguished by its adoption of a water-cooling Direct Liquid Cooling (DLC) system to accommodate the AI computing environment densely packed with high-performance GPUs. As demand rises for generative AI and large-scale AI model training and inference, data centers now require GPU servers with much higher computing power than in the past. Because high-performance GPUs consume substantial power and generate significant heat, efficient cooling technology is essential.


Looking up at the ceiling of the data center reveals water-cooling pipes running side by side. These pipes are filled with water, which absorbs and carries away the heat from the data center. Since AI data centers (AIDC) must accommodate more GPUs in limited space and deliver higher computational power, traditional air cooling alone cannot handle the increased thermal load. The weight of the water is also considerable, so AIDC designs must account for both equipment and water load to ensure structural safety.


The water-cooling DLC method uses a cold plate that directly contacts heat-generating parts, such as the GPU and Central Processing Unit (CPU), inside the servers—absorbing and removing heat. While conventional air cooling dissipates equipment heat using airflow, water cooling uses liquids with a far higher thermal conductivity to address high-density and high-heat environments. For example, a B200 requires 14.3 kW of power in an air-cooled system, but with water cooling, this drops to 12.3 kW. Based on NHN Cloud’s operation of AIDCs in Gwangju and Seoul, the rate of system faults was lower with water cooling than with air cooling. Noise levels were also much lower with water cooling, enabling conversations even in corridors lined with populated server racks.



[Report] Inside NHN Factory X Seoul, the Largest Hub of Over 7,000 High-Performance GPUs View original image

In the central control room visited by the reporter, staff were on site, monitoring the major data center equipment and IT infrastructure 24 hours a day. During the day, three staff and at night, two staff monitor GPU server status, cluster operations, and network connectivity in real time. Lee Iljun, Director of Technology Operations at NHN Cloud, said, "AI data centers are designed with sophisticated power, cooling, network, and server architecture to support AI training and inference workloads that utilize large-scale GPUs, making monitoring systems critical. If coolant supply is disrupted or the temperature of the coolant rises excessively, GPU performance drops and, in some cases, the GPUs may even shut down automatically, which is why we pay close attention to monitoring the cooling water as well."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing