‘Operation efficiency at core of AI infrastructure capability’

SEJONG — Naver Cloud showcased its hyperscale information middle GAK Sejong, expressing its confidence and effectivity in designing and working information facilities, crucial infrastructure for synthetic intelligence (AI) companies.

Throughout a press occasion on the middle on Monday, Naver Cloud Chief Info Officer Lee Sang-jun stated the competitiveness of AI infrastructure relies upon not solely on what number of graphics processing models (GPUs) an organization secures, but additionally on how stably and effectively it operates these assets.

“By sustaining the stability between GPU sourcing and internalization of operational applied sciences, Naver Cloud is strengthening its AI infrastructure competitiveness,” he stated.

“Today, doing AI enterprise requires large funding: large-scale information facilities like this one, and the massive quantity of electrical energy they eat,” Lee added. “However increasing these issues doesn’t occur in a single day. In contrast to world tech giants that spend on an unimaginable scale, we have now to pursue essentially the most environment friendly method attainable. I consider that by means of technological internalization, we are going to finally attain that degree.”

This undated handout picture reveals Naver’s GAK Sejong information middle in Sejong Metropolis. The bigger constructing above is the north wing of the ability, whereas the smaller construction related to it serves as the principle constructing. Courtesy of Naver Cloud

Occupying a 294,000-square-meter plot within the west-central area of Sejong Metropolis, GAK Sejong is portal and cellular large Naver’s second proprietary information middle, extensively considered as Korea’s first information middle devoted to empowering AI companies. GAK is known as after Janggyeonggak, referring to a constructing the place Buddhist or Confucian scriptures are saved.

GAK Sejong started operations in November 2023 as a high-density GPU computing facility able to dealing with each AI coaching and inference concurrently. GAK Sejong is supplied to accommodate as much as 600,000 server models upon the completion of its growth section, offering a complete storage capability of 65 exabytes.

GAK Sejong at the moment consists of a predominant constructing and a accomplished North Wing, with plans underway to assemble a South Wing within the coming years. The North Wing will bear three growth phases, adopted by phases 4 to 6 within the South Wing. Presently, solely the primary section is operational, whereas the second and third phases are scheduled for completion in 2027 and 2029.

Naver Cloud Chief Info Officer Lee Sang-jun speaks throughout a press convention at Naver’s GAK Sejong information middle in Sejong Metropolis, Monday. Courtesy of Naver Cloud

To deal with the big quantity of electrical energy consumed by AI accelerators, GAK Sejong is situated simply 2 kilometers away from a substation, designed to provide as much as 275 megawatts of energy. There are additionally plans to make use of a cooling system that mixes direct air, oblique air and chilled water to maintain server temperatures at a desired degree between 22 and 26 levels Celsius.

This cooling system embodies Naver’s gathered experience in technological internalization.

The primary-phase server rooms at the moment use an air cooling system primarily based on Naver’s proprietary designs. The corporate developed its first air misting unit in 2013 for the GAK Chuncheon information middle, adopted by the Naver Air Membrane Unit (NAMU) I in 2014, NAMU II in 2017 and the most recent NAMU III, which is put in at GAK Sejong. NAMU III is a hybrid system that makes use of direct air cooling in spring and fall, and oblique air cooling throughout summer season and winter.

The inside of GAK Sejong’s NAMU III air cooling system / Courtesy of Naver Cloud

The at the moment working North Wing spans roughly 350 meters in size and consists of 5 tales — three underground and two aboveground. It was designed to face up to earthquakes of as much as magnitude 7 and makes use of exterior air for cooling round 240 days a 12 months to maximise vitality effectivity.

There are 4 server rooms within the first section, every able to housing as much as 600 servers. This reporter visited one in all them, the place Nvidia’s H100 GPUs are in operation. A monitoring panel displayed the real-time energy consumption of every server, and each partitions of the room had been geared up with NAMU III air cooling programs. The room maintained a pleasantly cool temperature with a faint breeze, although the noise degree was fairly excessive.

Naver’s GAK Sejong information middle head Roh Sang-min, left, speaks throughout a press convention on the information middle in Sejong Metropolis, Monday. Courtesy of Naver Cloud

Roh Sang-min, head of GAK Sejong, stated that the second and third growth phases will introduce direct liquid cooling (DLC), which is required by future servers akin to Nvidia’s B200 and B300.

Together with DLC, Roh stated Naver can be conducting feasibility checks for immersion cooling, a way that submerges servers immediately in liquid, at GAK Chuncheon. Whereas Nvidia’s present AI servers don’t help immersion cooling, the U.S. tech large plans to introduce it beginning with next-generation platforms following the Rubin Extremely sequence.

With one-sixth of GAK Sejong’s deliberate full capability at the moment in operation, Roh stated the ability’s annual electrical energy prices have already reached about 22 billion gained ($15.34 million), and the quantity is anticipated to rise sharply within the coming years, requiring Naver to proceed investing assets into information middle operations.

Autonomous mobility robotic SeRo, left, and GaRo, proper, function in an IT storage room at Naver’s GAK Sejong information middle in Sejong Metropolis on this undated handout picture. Courtesy of Naver Cloud

“As we enter the age of AI information facilities, the facility density has elevated to a very completely different degree in comparison with after we used CPUs up to now,” Roh stated. “Lots of of GPUs function collectively as a single cluster, consuming a extremely concentrated quantity of electrical energy, and server rooms are already utilizing area rather more compactly.”

“From the second half of this 12 months by means of subsequent 12 months, home information facilities will proceed to face a scarcity of obtainable area. Beginning in 2027, extra capability will likely be accessible, however demand will doubtless stay unpredictably excessive. We opened GAK Sejong in November 2023, and it’s already full, so we’re finishing up each the second and third growth phases concurrently.”

The primary management room of Naver’s GAK Sejong information middle in Sejong Metropolis / Courtesy of Naver Cloud

CIO Lee stated that contemplating rising demand, Naver Cloud’s GPU-as-a-service (GPUaaS) enterprise is displaying sturdy development.

“Naver Cloud will leverage its expertise in AI infrastructure operations to increase the GPUaaS mannequin and create an ecosystem the place corporations can simply undertake AI,” he stated. “By way of this, we intention to make AI infrastructure not only a company asset however a basis for development throughout industries.”