Samsung seeks to regain memory leadership with low-power module

Samsung Electronics' SOCAMM2 / Captured from Samsung Electronics official website

Samsung Electronics is seeking to regain its leadership in the memory market with its latest low-power server memory module, the small outline compression attached memory module 2 (SOCAMM2), by supplying samples to a major customer as part of the final stage before mass production.

Samsung Electronics recently revealed that it began supplying its SOCAMM2 samples to a major client company. The company did not specify the client or details of the validation progress, but industry officials assume the samples are customer samples for Nvidia’s next-generation Vera Rubin artificial intelligence (AI) platforms.

Chip samples are typically categorized into three stages: working die, engineering sample and customer sample. A working die refers to the earliest functional stage of a chip. An engineering sample is an early preproduction version sent to customers to verify compatibility with their systems and products. After reflecting customer feedback, a near-final version known as a customer sample is delivered for qualification. Once the qualification process is completed, the chip goes into mass production.

SOCAMM2 is a module of low-power double data rate (LPDDR) memory chips, aimed at cutting power consumption to roughly one-third of conventional DDR-based modules. According to a Shinhan Securities report, applying the latest LPDDR5X to servers can reduce power consumption by about 77 percent compared with DDR5, while increasing bandwidth by 35 percent.

While high-bandwidth memory (HBM) chips are usually mounted on the graphics processing unit (GPU) for faster data processing, SOCAMM2 is usually mounted next to the CPU and focuses on handling large data volumes with greater power efficiency. In this setup, HBM supports computing acceleration, while SOCAMM2 is responsible for improving overall system-level power efficiency.

The semiconductor industry is now working to set up a standard for SOCAMM2, and an increasing number of chip developers are paying attention to the module’s power efficiency as AI service firms scramble to reduce the enormous amount of electricity consumed by data centers.

According to Samsung, its 192-gigabyte SOCAMM2 is 57 percent smaller than conventional DDR-based server modules and delivers more than twice the bandwidth while consuming at least 55 percent less power. The company noted that this enables stable performance even under intense AI workloads and offers high design flexibility, allowing it to be widely deployed across a range of AI systems.

Unlike traditional soldered LPDDR solutions, SOCAMM2 enables easy memory upgrades or replacements without any mainboard modification, helping system administrators minimize downtime and dramatically reduce the total cost of ownership.

To meet growing market demand for low-power memory, Samsung Electronics said it is seeking to expand the ecosystem for LPDDR-based server memory, particularly through collaboration with Nvidia to optimize SOCAMM2 for Nvidia’s acceleration infrastructure.

Along with Samsung, memory giants SK hynix and Micron are also scrambling to supply SOCAMM2 for Nvidia.

SK hynix is reportedly undergoing sample validation for its 192-gigabyte SOCAMM2 product at Nvidia, while Micron has already delivered its 192-gigabyte SOCAMM2 to the market. However, industry officials said Samsung has secured the largest share in the Nvidia-bound supply so far, due to the company’s technical collaboration with Nvidia and its superior supply capacity compared with its rivals.

“As the era of AI inference accelerates, system performance is likely to be determined not by the GPU’s raw computing power but by how memory bottlenecks are addressed,” an industry official said. “As leading cloud service providers are now prioritizing power efficiency for CPUs and memories, SOCAMM2 can be seen as the starting point of a new AI memory cycle.”