GPU memory system bandwidth improvement technology developed utilizing heterogenous memory and optical network

(From left) Professor Chung Myung-soo and Doctor Zhang Jie of Electrical Engineering Department [Photo provided = KAIST]

While being exclusively developed by global enterprises a Korean research team has developed an improved Graphic Processing Unit memory system technology by utilizing heterogenous memory system and optical network.

On the 2nd day of the month the research team lead by Professor Chung Myung-soo of KAIST Electronic Engineering Department announced that they improved performance of the system by 181% compared to the conventional system by developing the Ohm-GPU system using heterogenous memory composed of 3DXPoint and DRAM memories communicated by optical network.

The conventional GPU shows high calculation speed since it is consisted of multiple arithmetic units but it cannot show its full performance due to its low memory capacity of single DRAM and low data transmission bandwidth. The capacity may be improved by substituting it to 3DXpoint memory but the read or write speed will significantly drop. Also HBM may be used instead for higher bandwidth while it has limited copper wires installable to realize high bandwidth.

The research team enabled higher capacity with same performance by integrating the two memory systems. Also it improved the bandwidth by using single optical channel and overcame the limitations of the conventional GPU memory system.

Ohm-GPU processes all the memory with optical signals by manipulating the memory controller and interface. Usually the process is made in DRAM cache memory but some absent memories must be read from 3DXpoint. The overhead between two memories are mitigated by setting the wavelength of data movement and memory access differently and using 3DXPoint controller in data movement by minimizing interference of the memory controller.

The newly developed GPU technology showed 181% improved performance in graph processing and science application. The technology is expected to substitute high performance accelerator memory system used in areas like AI big data cloud computing etc.

Chung said that GPU memory system is lead by some distinguished enterprises and the research somehow opened possibilities to superiority in GPU and similar memory system.

The research will be presented in International Symposium on Microarchitecture (MICRO) 2021 this October in the name of Ohm-GPU: Integrating New Optical Network and Heterogeneous Memory into GPU Multi-Processors. The research team is sharing the result in all 4 big academic symposiums in 2021. The details of the research can be found in the lab website (http://camelab.org).

Website: : http://cms.wip-news.com/news/articleView.html?idxno=8431