SK Hynix Memory Ramp

SK hynix said on April 20 that it has started mass production of a 192-gigabyte server memory module called SOCAMM2 for artificial intelligence systems. (upi.com) The company said the module is built on its 1c nanometer-class dynamic random-access memory process and is aimed at Nvidia’s next Vera Rubin platform. SK hynix said SOCAMM2 delivers more than twice the bandwidth of a conventional registered dual in-line memory module, or RDIMM, with more than 75% better power efficiency. (tmcnet.com) SOCAMM is a low-power server memory format for AI machines that need to move large amounts of data between processors and memory. SK hynix showed SOCAMM2 at CES in January and then moved it into volume manufacturing in April. (news.skhynix.com; upi.com) The shift from trade-show demos to mass production comes as cloud companies and chipmakers are building larger AI servers around Nvidia’s next graphics processing unit roadmap. SK hynix said in March that it had reaffirmed its partnership with Nvidia at GTC 2026 and was presenting an expanded AI memory lineup there. (news.skhynix.com; news.skhynix.com) SK hynix has spent the past year moving multiple AI memory products from sampling into manufacturing. In September 2025, the company said it had completed development of HBM4 and finished preparations for mass production after shipping 12-layer HBM4 samples to customers earlier. (news.skhynix.com; news.skhynix.com) High bandwidth memory, or HBM, is the stacked memory used next to AI processors to feed them data quickly. SK hynix’s HBM4 page says the product can deliver more than 2.8 terabytes per second of bandwidth and up to 16-layer stacking. (product.skhynix.com) The company has also been ramping the current generation. In September 2024, SK hynix said it began volume production of 12-layer HBM3E with 36 gigabytes capacity and 9.6 gigabits-per-second speed, then planned shipments to customers within the year. (news.skhynix.com) SK hynix said in its January 2026 market outlook that demand for HBM3E and HBM4 would drive an “HBM-led memory supercycle” as AI infrastructure spending expands. The new SOCAMM2 ramp puts another one of those AI memory bets into factory output. (news.skhynix.com; upi.com)