(SK hynix)
SK hynix Inc. announced on Monday that it has begun mass production of the 192 gigabyte (GB) small outline compression attached memory module (SOCAMM2), a low-power memory module for servers based on LPDDR5X, signaling a broader low-power chip race.
According to SK hynix, it has started full-scale mass production of a 192GB SOCAMM2 product based on its sixth-generation 10-nanometer-class (1c) LPDDR5X DRAM.
SOCAMM2 manufactured using the 1c process achieves more than double the bandwidth and over 75 percent improved energy efficiency compared to existing Registered Dual In-line Memory Modules (RDIMM) module technology.
It is also designed to be optimized for Nvidia Corp.’s next-generation AI accelerator, Vera Rubin, which is set for release in the second half of this year.
Ahead of SK hynix, Samsung Electronics Co. began mass production of SOCAMM2 for Nvidia supply in March. Its product uses fifth-generation 10-nanometer-class (1b) DRAM, and SK hynix’s version is expected to offer higher performance. However, industry observers expect Samsung Electronics to lead in overall supply volume to Nvidia, having started shipments earlier.
Micron Technology Inc., which began developing early SOCAMM models with Nvidia, introduced a 256GB product in March and has supplied samples to customers. SOCAMM2 is also referred to as a second HBM due to its role as a high-performance memory product for AI servers.
Rising demand for AI servers has significantly boosted demand across the memory semiconductor sector.
As global growth is increasingly fueled by investment in AI servers and infrastructure, SOCAMM is emerging as a key driver for Samsung Electronics and SK hynix alongside high bandwidth memory (HBM). Advanced Micro Devices Inc. and Qualcomm Inc. are also reviewing adoption of SOCAMM in their servers, pointing to further market expansion.
SOCAMM2 uses low-power LPDDR commonly found in smartphones, which is why it is also cited as a factor behind recent increases in LPDDR prices for mobile devices. The 192GB SOCAMM2 “sets a new standard for AI memory performance,” the company said.
As inference workloads gain importance in AI systems, demand is rising for a broader mix of memory types in AI servers. SOCAMM has emerged as a module positioned between high bandwidth memory (HBM) and DDR5, addressing different performance and power needs.
DDR5 remains the standard general-purpose memory used in PCs and servers, typically deployed in RDIMM configurations for scalable, high-capacity systems.
As memory architectures diversify across HBM, DDR5 and SOCAMM, chipmakers are shifting from mass production of standardized products toward customized, integrated memory solutions tailored to specific workloads.
댓글
0