NVIDIA and the Korea Advanced Institute of Science and Technology (KAIST) established a joint AI research lab focused on synthetic data generation and learning efficiency, with Director Kim Hyun-woo emphasizing in a media interview on the 10th that quality data and efficient learning techniques are core to AI competitiveness. The partnership, announced in July, involves a 5-year collaboration worth a total of $300 million (approximately 403.7 billion KRW), with $250 million (approximately 336.4 billion KRW) allocated to computing resource support. NVIDIA selected KAIST as a research partner based on the institution's global AI research capabilities, highlighted by KAIST's ranking as the world's second institution by first-author papers at the Neural Information Processing Systems (NeurIPS) conference with 166 accepted papers, surpassing MIT and Stanford.
NVIDIA-KAIST Partnership Allocates $250 Million for Computing Resources
The joint research lab provides annual research funding and NVIDIA internship opportunities to a minimum of 10 KAIST researchers. Separately, the lab plans to hire full-time NVIDIA-affiliated researchers. Unlike traditional industry-academia collaborations, this partnership enables domestic researchers to conduct joint research with NVIDIA's global research teams and utilize large-scale GPU resources. Director Kim stated that researcher hiring has begun, with plans to select some personnel within this year. The total research workforce size and GPU procurement quantities have not yet been finalized. Kim described NVIDIA's computing resource support as an "opportunity like timely rain" for domestic researchers facing GPU supply difficulties.
KAIST Ranks Second Globally in NeurIPS 2024 with 166 Papers
KAIST achieved the world's second position in research output at NeurIPS based on first-author institutional affiliation, with 166 accepted papers trailing only China's Tsinghua University. This ranking places KAIST ahead of the Massachusetts Institute of Technology (MIT) and Stanford University. Director Kim stated that KAIST has become "a school that the world pays great attention to," highlighting the institution's world-class AI research competitiveness. Kim assessed that the collaboration establishes a "global fast track" for Korean researchers to directly enter the global AI research ecosystem through exchanges with NVIDIA's global research teams and access to large-scale computing resources.
Research Lab Addresses Korean Language Data Scarcity Through Synthetic Generation
The lab's core research directions focus on generating quality synthetic data and improving AI model learning performance by utilizing existing data more efficiently. Director Kim pointed out that while domestic AI industry has concentrated on computing infrastructure like GPUs and model development, attention to securing quality data and learning methods has been relatively insufficient. In the multilingual training dataset mC4, English data comprises approximately 3.93 billion entries while Korean data contains approximately 24 million entries, with English exceeding Korean by approximately 163 times. Kim emphasized the need for technologies that use AI to generate new synthetic data and learn more from limited data, given the difficulty of securing Korean language data at the same scale as English.
LG AI Research Applies Nemotron Dataset in EXAONE 2.5 Development
LG AI Research's "EXAONE 2.5," released on the 9th, utilized NVIDIA's Nemotron dataset including "Nemotron-Personas-Korea" for training. Nemotron-Personas-Korea is an open dataset that synthesized 7 million virtual persona data reflecting Korean demographic characteristics. According to Director Kim, the dataset ranked first among popular datasets on the global AI platform Hugging Face for 10 days and recorded approximately 140,000 downloads. While not a direct research outcome of the joint lab, this demonstrates the potential application of open data research that the lab will pursue. The lab plans to develop and release data generation and learning methodologies applicable not only to specific NVIDIA models but also to AI models from other companies. Kim stated, "If what we produce can be used elsewhere in our country, we consider that the success of our research."
Lab Plans Full-Time Researcher Hiring with Results Expected in Second Half of Next Year
The research lab is currently refining research directions while proceeding with personnel recruitment and computing infrastructure establishment. Director Kim stated that full-time researcher hiring has started with plans to select some personnel within this year. Kim declined to disclose specific research projects but expected major outcomes to be released around the second half of next year. Kim anticipated that the joint lab will also serve as an opportunity to attract Korean AI talent active overseas back to Korea, stating, "Bringing our country's talent from overseas back home is one of our major goals" and emphasizing that "this joint lab confirms KAIST's outstanding research capabilities to the entire world." Director Kim received a computer science doctorate from Seoul National University and conducted research at the Allen Institute for Artificial Intelligence (Ai2) and NVIDIA in the United States. He joined KAIST's Kim Jaechul Graduate School of AI as an assistant professor in August and currently leads the joint AI lab while serving concurrently as an NVIDIA research manager.
FAQ
What is the scale of the NVIDIA-KAIST joint AI research lab partnership?
The partnership involves a 5-year collaboration worth a total of $300 million (approximately 403.7 billion KRW), with $250 million (approximately 336.4 billion KRW) allocated to computing resource support. The lab provides annual research funding and NVIDIA internship opportunities to a minimum of 10 KAIST researchers.
Why did NVIDIA select KAIST as a research partner?
NVIDIA selected KAIST based on the institution's global AI research capabilities. KAIST ranked second globally in first-author papers at the Neural Information Processing Systems (NeurIPS) conference with 166 accepted papers, surpassing MIT and Stanford University and trailing only China's Tsinghua University.
