KAIST develops tech that use optical network to boost GPU memory performance

A South Korean research team has developed a technology that uses optical networks to boost GPU capacity and bandwidth.

A team led by electrical engineering professor Myoungsoo Jung of Korea Advanced Institute of Science and Technology (KAIST) combined a DRAM and 3D Xpoint memory for a heterogeneous memory system.

The team developed an Ohm-GPU technology that has this system communicate with an optical network.

Ohm-GPU adjusts the memory controller and interface inside the GPU to process all memory request from the heterogeneous memory system into optical signals.

Jung’s team also minimized the latency when data are transferred between the heterogeneous memories.

Use of Ohm-GPU increased the memory performance of the system by 181% compared to just using DRAM, KAIST said.

The team said it utilized optical network’s ability to use a single optical channel to transfer multiple optical signals with different wavelengths, according to KAIST.

The research paper on the technology will be introduced at Micro 2021 symposium in October.