Chip giants, opening up a new battlefield.

Samsung Electronics and SK Hynix have respectively partnered with US chip giants Nvidia and Intel, intensifying competition in the low-power server memory field. Both companies have released their own products: SK Hynix has introduced a new high-capacity, high-performance general Dynamic Random Access Memory (DRAM) product, while Samsung has launched a new generation of memory modules. The rise of artificial intelligence has driven a surge in memory demand, leading to an increase in prices for High Bandwidth Memory (HBM) and traditional server DRAM. This has resulted in increasingly fierce competition between Samsung and SK Hynix as both companies vie for dominance in the DRAM market. SK Hynix announced on Thursday that its 256GB DDR5 Registered Dual In-Line Memory Module (RDIMM) based on 32Gb fifth-generation 10-nanometer-class (1b) DRAM had passed Intel's compatibility and performance validation. RDIMM is a type of memory module designed to ensure system stability even with multiple DRAM chips installed. High-capacity products are primarily used in enterprise data centers. "This product is currently the highest-capacity RDIMM on the market based on 1b 32Gb DRAM, and the first RDIMM of the same specifications to be certified by Intel," SK Hynix stated. While Intel lags behind Nvidia in the GPU field, it still holds over 70% of the market share in the data center CPU market. SK Hynix's RDIMM has now been certified and can be immediately applied to servers based on Intel's latest Xeon 6 CPU. Once Intel has verified the performance and compatibility of memory products, server manufacturers such as Dell, HP, and Supermicro typically make bulk purchases for server production. SK Hynix's new RDIMM memory utilizes Through-Silicon Via (TSV) technology, which is the core technology of HBM. By stacking 32Gb DRAM chips into a 3D structure and connecting them through micro vertical channels, SK Hynix has achieved an industry-leading capacity of 256GB. Developers at SK Hynix stated, "Compared to existing products, servers using this product can achieve a 16% increase in AI inference performance and an 18% reduction in power consumption." Meanwhile, Samsung Electronics announced on Thursday that they had completed the development of a memory product optimized for AI data centers - the Small Compressible Add-On Memory Module (Socamm2) - and had provided samples to Nvidia. Socamm2 is a server memory module consisting of stacked low-power Double Data Rate DRAM. Compared to common RDIMMs, Socamm2 has a smaller size, more than doubled bandwidth, and a 55% reduction in power consumption, making it ideal for high-performance AI computing. Unlike HBM packaged in GPUs, Socamm2 is a detachable module, making the manufacturing process simpler and providing greater design flexibility. Due to unresolved issues with international standardization and platform compatibility, Socamm2 has not yet been commercialized. It is reported that Nvidia plans to use Socamm2 in its next-generation AI platform Vera Rubin, which is expected to be released in the second half of next year. Samsung Electronics stated, "We are working with global major partners to lead the establishment of international standards for Socamm2." Dion Harris, Senior Director of High-Performance Computing and AI Infrastructure Solutions at Nvidia, said, "Through our continued collaboration with Samsung Electronics, we are constantly optimizing the next generation of memory such as Socamm2." All three major memory chip manufacturers are competing to supply Socamm2 to Nvidia. In October of this year, Micron, the world's third-largest memory chip manufacturer based in the US, officially announced that they had provided samples of Socamm2 to Nvidia. On the same day, Samsung Electronics and SK Hynix unveiled physical samples and specifications of their respective Socamm2 at a semiconductor expo in South Korea. This article was selected from the "Semiconductor Industry Observation" WeChat official account; edited by GMTEight: He Yucheng.