KAIST develops high-performance AI accelerator technology

[Courtesy of KAIST]

A South Korean research team has developed a high-performance graphic processing unit (GPU) technology for artificial intelligence (AI) accelerators utilizing computer express link (CXL), a next-generation interface technology.

According to the Korea Advanced Institute of Science and Technology (KAIST) on Monday, a research team led by Professor Jung Myoung-soo has developed the high-performance technology.

GPU-based AI accelerators are system semiconductors designed to enhance the AI computing capabilities of specific software.

CXL is an advanced interface that facilitates direct communication between the central processing unit (CPU), the accelerators, and storage devices.

The traditional method of memory expansion, which involves connecting multiple GPUs, faces the issue of rapidly escalating deployment costs due to the high price of GPUs.

The industry has been exploring a technology that connects large memory directly to GPU devices using CXL technology.

However, the performance of memory expansion devices directly connected to GPUs has been limited.

The KAIST team offered a solution by enabling the memory expansion device to determine the timing of memory writes, allowing simultaneous writing to GPU local memory without waiting for the GPU device to complete the memory write operation.

Its findings will be published at the USENIX conference in Santa Clara, the United States.