KAIST Opens Era of Terabyte Memory

- Professor Jeong Myungsoo of the Department of Electrical and Electronic Engineering Develops Large-Capacity, Persistent Memory Technology Combining Non-Volatile Memory (NVDIMM) with Ultra-Low Latency SSD
- Paper Accepted at ISCA, the Most Prestigious of the World's Top 3 Computer Architecture Conferences
- Opens Possibility for Domestic Technological Superiority in Persistent Memory Technology Previously Led Only by Some Overseas Companies

Overview of Terabyte Memory Technology Proposal. Photo by Korea Advanced Institute of Science and Technology (KAIST)

[Asia Economy Reporter Kim Bong-su] A memory technology that integrates non-volatile memory (NVDIMM) and ultra-low latency SSD (semiconductor storage device) into a single memory, significantly improving performance and capacity compared to existing technologies, has been developed by a domestic research team.

The Korea Advanced Institute of Science and Technology (KAIST) announced on the 16th that Professor Myungsoo Jung's research team from the Department of Electrical Engineering successfully developed Memory-over-Storage (MoS) technology, which integrates non-volatile memory and ultra-low latency SSD into one memory space. This technology reuses existing storage technology to provide storage capacity at the terabyte (TB=1024GB) level, which is more than four times that of Intel Optane per memory slot, while boasting user-level data processing speeds similar to volatile memory (DRAM).

Non-volatile memory refers to memory that adds flash memory and super capacitors to existing DRAM to retain data even during power outages, and ultra-low latency SSD is an improved SSD with very low latency.

Existing NVDIMM has the advantage that the CPU can directly access non-volatile memory without the help of the operating system, but it cannot handle large amounts of data because it uses DRAM as is and cannot infinitely increase battery size. Alternatives include Intel Optane memory (Intel Optane DC PMM) and Memory Drive Technology, but these require operating system assistance every time non-volatile memory is accessed, resulting in read and write speeds about 50% lower than NVDIMM.

The MoS technology proposed by Professor Jung’s team uses ultra-low latency SSD as the main memory and NVDIMM as cache memory, allowing users to use the large storage space of SSD as memory. At the same time, it achieves performance similar to that of using NVDIMM alone, fully overcoming the limitations of future persistent memory technologies.

MoS technology is applied to the memory controller hub (MCH) located on the motherboard or inside the CPU to handle all user memory requests. User requests are generally processed in the NVDIMM cache memory, but data not stored in NVDIMM must be read from the ultra-low latency SSD. While existing technologies rely on the operating system to handle these SSD reads, the developed MoS technology mitigates the input/output overhead of the operating system by having hardware inside the MCH directly handle SSD input/output.

Compared to software-based memory drives or Optane persistent memory technologies, it achieves a 45% reduction in energy consumption and a 110% improvement in data read/write speeds. This is expected to replace existing memory and future persistent memory used in data centers and supercomputers that require large memory capacity and are sensitive to system failures caused by power outages.

Professor Myungsoo Jung explained, "Although some leading overseas companies are driving future persistent memory technologies, this research achievement is meaningful in that it opens the possibility of gaining a competitive edge in the related market through domestic technology and existing storage and memory technologies."

The research results will be presented as a paper at ISCA 2021 (International Symposium on Computer Architecture), the top academic conference in the field of computer architecture, to be held in June this year.

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