Low-power dynamic random-access memory (LPDDR) is emerging in the artificial intelligence (AI) memory market. It is gaining attention as an alternative, promising to address AI bottlenecks that high-bandwidth memory (HBM) alone cannot solve, while also offering higher energy efficiency. 

Samsung Electronics and SK Hynix have already released next-generation LPDDR products, joining the competition for market share.

LPDDR6 R&D Achievements

Industry insiders revealed on February 24th that Samsung Electronics and SK Hynix announced their LPDDR6 development achievements at the ISSCC (International Solid-State Circuits Conference) 2026, held in San Francisco, USA, from February 15th to 19th (local time). ISSCC is considered the most prestigious conference in the semiconductor design field.

Both Samsung Electronics and SK Hynix showcased LPDDR6 memory compliant with JEDEC (Joint Electron Device Engineering Committee) standards, achieving a maximum transfer speed of 14.4Gb/s. This represents an approximately 35% speed improvement compared to the previous generation LPDDR5X (maximum transfer speed 10.7Gb/s). LPDDR6 is expected to be available in the second half of this year.

Samsung Electronics emphasized improved energy efficiency. Most of the circuitry in the company's design operates normally even at ultra-low voltages, resulting in a 27% reduction in read power consumption compared to LPDDR5. At the lowest voltage (0.97V), it achieved a transfer speed of 12.8Gb/s and enhanced stability in real-world environments through structural optimization.

SK Hynix's advantages

SK Hynix also employed an "efficiency mode" to reduce power consumption. The company reduced power consumption while maintaining a 12.8Gb/s transfer speed. SK Hynix stated that the product can provide a transfer speed of 10.9Gb/s in low-voltage environments and maintains stable performance even at the highest transfer speed (14.4Gb/s). Furthermore, the company demonstrated power control technology that uses only the required power.

Traditionally, LPDDR was primarily used in energy-intensive mobile devices such as smartphones. However, with the proliferation of artificial intelligence infrastructure, its application scope has expanded in recent years. LPDDR has attracted significant attention because it can compensate for the capacity limitations of single HBM memory modules while offering excellent energy efficiency. Using LPDDR can expand system-level memory capacity, improve data center energy efficiency, and reduce total cost of ownership (TCO).

Specifically, inference AI generates temporary data called a "key-value cache" (KV cache) to maintain the context of a conversation. As the context length increases, the KV cache capacity grows rapidly, which can lead to insufficient HBM memory capacity installed in the GPU (graphics processing unit), creating a bottleneck. Therefore, a design scheme that combines the high bandwidth of HBM with the high capacity and low power consumption of LPDDR is currently being explored.

The advantages of LPDDR

Furthermore, compared to standard DRAM, LPDDR offers higher power consumption and space efficiency, enabling the expansion of memory capacity without significant modifications to existing system architectures. Samsung Electronics and SK Hynix have also improved data management efficiency by integrating a "metadata" area into LPDDR6, thereby enhancing the computational efficiency of artificial intelligence.

For example, NVIDIA's GH200 system connects LPDDR5X memory installed in the Grace CPU (Central Processing Unit) with HBM3E video memory installed in the Hopper GPU (Graphics Processing Unit). NVIDIA's Vera CPU, planned for release in the second half of this year, is expected to be equipped with 1.5TB (terabytes) of LPDDR5X memory.

According to Micron Technology's research, expanding LPDDR5X memory capacity from 512GB to 1.5TB can reduce the first token response time (TTFT) by up to 98% in real-time inference environments. With increased memory capacity, the number of recalculations can be reduced, thus improving initial response speed.

An industry insider said, "HBM alone cannot meet all the rapidly growing memory demand," he added, "as the demand for server DRAM and LPDDR5X grows in tandem with the demand for HBM, this will improve the overall profitability of related memory products."

Source: Compiled from businesskorea

View more at EASELINK