The semiconductor industry, once focused on speed and capacity, is now shifting towards power efficiency. As AI learning and training become more advanced, the power required to process massive data has surged, leading to fierce competition to develop low-power chips.
AI chips, often dubbed “energy guzzlers,” are big electricity consumers. Nvidia’s upcoming high-performance AI chip, the B100, requires 1000 watts of power. The previous AI chip models, the A100 and H100, require 400W and 700W, respectively, indicating that higher performance leads to higher power consumption.
Samsung Electronics and SK Hynix, leading chipmakers that supply memory chips used in AI chips, are introducing new low-power semiconductor solutions. Semiconductor materials companies are also developing next-generation materials, such as glass, to improve power efficiency.
Low-power chips are crucial for devices like smartphones, tablets, and laptops, which need to conserve battery life while performing AI computations without being connected to the internet. “Low-power semiconductors can extend battery life in mobile devices and reduce the energy required for data processing in servers,” said a semiconductor industry insider. “As power supplies become scarcer, low-power semiconductors will become even more important.”
LPDDR, or lowpower double data rate memory, is at the forefront of low-power chips. Unlike conventional DRAM with a single data path, DDR has two paths, enabling faster data processing. LPDDR also reduces power consumption and has been developed up to the 7th generation (5X). These are commonly used in high-performance smartphones and laptops.
Both Samsung Electronics and SK Hynix are boosting LPDDR research and development. In April, Samsung developed LPDDR5X (7th generation), a low-power DRAM chip with the fastest data processing speed up to date. The company has recently completed speed verification and is reportedly preparing for mass production. This new product offers over 30% more capacity than the previous generation while reducing power consumption by 25%. The company has applied new technology that adjusts power based on performance and speed.
SK Hynix first commercialized LPDDR5T DRAM with a 5X performance upgrade late last year. This product is used in Chinese smartphone maker Vivo’s flagship model. It can process 15 full HD movies in one second while significantly reducing power consumption.
Recently, there has also been a push to develop LPDDR stacking technology. Like HBM (high-bandwidth memory), which stacks multiple DRAMs, LPDDR stacking aims to increase capacity and speed while minimizing power use.
Efforts to develop next-generation materials to enhance semiconductor power efficiency are also underway. Glass substrates, known as the “dream substrate” of the AI era, are a prime example. While current substrates use plastic or silicon to connect multiple semiconductors, smooth glass could dramatically increase data processing speed without increasing power consumption. Although no commercial products are available yet, the competition to capture this market is intense, according to industry insiders.
SKC’s subsidiary Absolix recently completed building a glass substrate factory in Georgia; the U.S. Samsung Electronics plans to build a prototype production line at its Sejong plant this year and begin mass production by 2026. LG Innotek officially launched its glass substrate business in March, with CEO Moon Hyuk-soo directly overseeing the initiative.
Low-power, high-performance chips based on gallium nitride (GaN) and silicon carbide (SiC) are being developed as potential replacements for traditional silicon. Samsung Electronics has set up a dedicated business team for GaN-based semiconductors and aims for mass production by 2025. “Until now, AI chips focused on securing HBM, but in the era of on-device AI, LPDDR is expected to become a core product,” said an industry insider. “Even Nvidia uses LPDDR DRAM rather than HBM for its CPUs.”
