KT "Building Large-Scale GPU Farm in Early 2022... Becoming a 'Helper' for Businesses"

Kim Juseong, Executive Director of KT Cloud/IDC Business Promotion Office
Explains at Online Press Conference

[Asia Economy Reporter Minyoung Cha] "We currently have GPU resources on the scale of 100 units, and we plan to expand this to a massive GPU farm with thousands of units by next year."

On the 27th, Kim Juseong, Executive Director of KT Cloud/IDC Business Promotion Office, stated this during an online press briefing while introducing KT's new product, 'Hyperscale AI Computing (HAC).' HAC, which KT unveiled this month, is a cloud-based GPU infrastructure service developed in collaboration with AI infrastructure solution startup 'More.' GPU infrastructure is essential for artificial intelligence (AI) companies but is a market monopolized by a few companies such as Nvidia. The cost burden is also significant. KT aims to resolve monopoly issues and reduce costs to lead the localization of GPU technology.

The biggest advantage of HAC is its reasonable cost structure. It is the first in Korea to offer a 'pay-as-you-go' system where users pay only for the GPU resources they use at their desired time and then return them. Executive Director Kim explained, "The cost reduction effect from using HAC is about 70%, and we plan to run a discount promotion until the first half of next year."

It also allows flexible responses to the scaling up of AI models by demand companies, such as in the case of ultra-large AI. This is because, with recent advancements in AI technology, the amount of model data training per company is increasing. For example, GPT-3 requires over 1,000 times more parameters. Kim said, "Existing GPU services had difficulties because development was done on a single GPU with small memory and then expanded to multi-GPU, requiring all existing data to be deleted and restarted from scratch. KT's HAC can allocate hundreds to thousands of large-scale GPU clusters, enabling effective response even in large-scale situations."

KT has also established a short-term advancement plan. The reference models currently provided by KT are divided into three major groups?vision, NLP, and others?with 16 models in total. Through monthly updates, this will increase to 41 models by January next year. KT will also expand beyond the current limitation of only providing PyTorch to include TensorFlow by next year. Executive Director Kim emphasized, "Our goal is to focus on the role of 'enabler' for companies and to achieve a certain market share through competition with Nvidia by 2023."

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