 |
Dongsoo Han, Professor of Electrical and Electronic Engineering, KAIST / Photo = KAIST |
[Alpha Biz= Reporter Kim Jisun] KAIST announced on the 19th that a research team led by Professor Han Dong-Soo of the School of Electrical Engineering has developed a technology that can accelerate AI model training by over 100 times, even in limited network environments. This breakthrough was achieved in collaboration with a research team from UC Irvine.
Traditionally, AI model training requires expensive high-performance server GPUs, such as Nvidia's H100, and high-speed networks with capacities up to 400Gbps. The cost of such infrastructure poses a significant challenge for small IT firms and academic research teams.
The research team developed a distributed training framework called "StellaTrain," which enables efficient AI training using consumer-grade GPUs, which are only a fraction of the price of the Nvidia H100, even in standard internet environments.
The slowdowns in AI training with cheaper GPUs are typically due to their limited memory and network bandwidth. The team overcame these issues by utilizing CPUs and GPUs in parallel, allowing them to divide and process tasks more efficiently. They also implemented a system where the amount of data transferred between GPUs could be dynamically adjusted based on network conditions, making it possible to achieve faster training speeds without the need for high-speed networks.
When the team applied the StellaTrain technology, it demonstrated performance improvements of up to 104 times compared to traditional methods.
Professor Han remarked, "This research will greatly contribute to making large-scale AI model training more accessible to everyone. We will continue to develop technologies that enable AI training in low-cost environments."
The research was presented at the "ACM SIGCOMM 2024" conference in Sydney, Australia, in August. It was supported by the Ministry of Science and ICT, IITP, and Samsung Electronics.
Alphabiz Reporter Kim Jisun(stockmk2020@alphabiz.co.kr)
