KAIST develops StellaTrain technology
Efficient learning with 20 times cheaper GPU, ‘[Seoul Economy]’,
,
, ‘Domestic researchers have developed a technology that can train artificial intelligence (AI) models cheaply without expensive infrastructures like high-performance graphics processing units (GPUs).’
,
,
StellaTrain technology diagram. Photo provided by KAIST,
,
, ‘The Korea Advanced Institute of Science and Technology (KAIST) announced on the 19th that a research team led by Professor Han Dong-soo of the Department of Electrical and Electronic Engineering has developed the ‘StellaTrain’ distributed learning framework that can accelerate AI model training by tens to hundreds of times using consumer-grade GPUs. The research results were presented at ‘ACM SIGCOMM 2024′ held in Sydney, Australia, last month.’,
,
, ‘Previously, high-performance GPUs costing thousands of dollars per unit were needed for AI model training. NVIDIA’s H100 is a typical example. This kind of infrastructure is a significant cost burden for many AI development companies except for a few big tech companies.’,
,
, ‘The research team has developed a technology that can efficiently train AI models using consumer-grade GPUs that are 10-20 times cheaper than the H100, even with a standard internet environment. Low-cost GPUs faced the problem of slow AI training speed due to limitations in memory capacity and network speed. The research team solved this problem by configuring multiple GPUs in parallel and developing customized algorithms. StellaTrain by the research team showed up to 104 times faster learning performance compared to existing methods.’,
,
, ‘Professor Han stated, “This research will contribute to making large-scale AI model training easily accessible to everyone” and said, “We plan to continue developing technologies that enable training large-scale AI models in low-cost environments in the future.”‘,
,