South Korean researchers advance AI computing and training efficiency

Research into improving artificial intelligence (AI) efficiency is accelerating, with a focus on moving data processing closer to memory to reduce latency and power consumption. At the MICRO 2026 international conference, papers regarding memory-related topics saw a significant increase, rising from 11 last year to 63 this year.

South Korean companies Samsung Electronics and SK hynix are actively developing next-generation memory technologies. Samsung is collaborating with Nvidia and various universities on technologies like AMMA, which distributes attention operations across chiplets to reduce latency. SK hynix, in partnership with KAIST, has developed NELSSA, a system that allocates tasks between GPUs and Processing near Memory (PNM) to increase AI response throughput by up to 5.5 times.

Separately, researchers at KAIST have developed a clustering technology to optimize AI training. By categorizing training data based on its educational effect and distributing it across multiple GPUs, the team reduced redundant learning. Testing showed that using only 31% of total training data could maintain performance while reducing training time by up to 69.6%.