"Called Price" Graphics Cards... Korea Secures Lead with GPUs Twice as Fast

Jeong Myungsoo KAIST Professor Team Presents Ohm-GPU Technology
Opens Possibility to Secure Market Dominance in High-Performance Accelerator Memory Systems

Not related to the article content. Photo by Asia Economy DB

[Asia Economy Reporter Kim Bong-su] A technology that can nearly double the performance of computer graphics cards, which have become "sold out due to the cryptocurrency mining craze," has been developed by a domestic research team.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 2nd that Professor Myungsoo Jung's research team from the Department of Electrical Engineering and Computer Science succeeded in developing the 'Ohm-GPU' technology, which communicates via an optical network in a heterogeneous memory system integrating 3D XPoint memory (XPoint) and DRAM memory, achieving a performance improvement of over 181% compared to existing systems.

Existing GPUs (Graphics Processing Units) consist of multiple computing units, making them very fast. However, due to the low memory capacity and narrow data transfer bandwidth of memory systems using only DRAM, their performance could not be fully utilized.

Replacing DRAM with XPoint is an option, which provides eight times larger memory capacity, but the read and write performance drops by four and six times, respectively. To increase bandwidth, HBM (High Bandwidth Memory), which stacks DRAM in 3D to achieve high bandwidth, can be used, but due to the limitation in the number of electrical channels (copper wires) that can be mounted within a single area, it is difficult to meet the high bandwidth required by GPU memory systems.

The Ohm-GPU technology developed by Professor Jung's team adopted a heterogeneous memory system integrating large-capacity XPoint and high-performance DRAM. It increased memory capacity while maintaining the same performance as existing memory systems. Additionally, it significantly expanded memory bandwidth by leveraging the advantages of optical networks, which can transmit multiple optical signals of different wavelengths through a single optical channel (optic fiber). This successfully addressed the limitations of existing GPU memory systems.

Ohm-GPU technology modifies the memory controller and interface inside the GPU to handle all memory requests of the heterogeneous memory as optical signals. Memory requests are generally processed in the DRAM cache memory, but data not present in DRAM must be read from XPoint. The overhead (latency) caused by data movement between heterogeneous memories is mitigated by setting different optical wavelengths for memory access for computation and for data movement. Furthermore, by minimizing memory controller intervention and having the XPoint controller perform data movement between heterogeneous memories, the overhead is alleviated.

The developed Ohm-GPU technology achieved a 181% performance improvement in various graphics processing and scientific application executions compared to existing electric network-based GPU memory systems using only DRAM. It is expected to replace memory systems of high-performance accelerators requiring large capacity and high bandwidth data transfer, such as artificial intelligence (AI), big data, and cloud computing.

Professor Jung emphasized, "GPU memory system technology is currently led by some leading overseas companies," adding, "This research achievement is significant in that it opens the possibility of gaining a competitive edge in the market related to GPU and all high-performance accelerator memory systems similar to GPUs."

This content was produced with the assistance of AI translation services.