Samsung is reportedly in secret talks with NVIDIA to develop a new type of advanced memory chip for AI processors. Known as SOCAMM, or system on chip advanced memory module, the new memory standard is being hailed as the second high-bandwidth memory (HBM). It could significantly reshape the memory semiconductor landscape in the AI era. NVIDIA is also discussing this project with SK Hynix.
Samsung might work with NVIDIA on a new AI memory chip
NVIDIA is the world’s largest maker of AI processors. South Korean semiconductor firm SK Hynix supplies the bulk of HBM for these accelerators. Its compatriot Samsung is still unable to achieve NVIDIA’s approval for its HBM3E chips, though it might be close to finally meeting all requirements. In the meantime, the American giant is discussing the potential commercialization of SOCAMM with both Samsung and SK Hynix.
SOCAMM, short for System on Chip Advanced Memory Module, is an innovative DRAM module designed to boost the performance of AI supercomputers. It has a superior cost-performance ratio, using LPDDR5X DRAM for low power consumption and high efficiency. With 694 I/O ports — 50 more than the existing LPCAMM standard — SOCAMM reduces data bottlenecks, improving AI computing.
Additionally, its compact size allows more DRAM modules in the same space, enhancing performance further. Unlike traditional soldered memory, SOCAMM is detachable, making upgrades and replacements easy. This flexibility is particularly beneficial in AI computing, where evolving workloads demand scalable and efficient memory solutions. In short, SOCAMM is the way forward to bring AI PCs to the masses.
NVIDIA is pursuing its own memory standard
Unlike conventional memory standards established by the Joint Electron Device Engineering Council (JEDEC) through industry-wide consensus, NVIDIA is taking an independent approach by spearheading its own memory standard. This reflects its market dominance and strategic direction, as well as its confidence. The firm wants to sign Samsung and SK Hynix as key manufacturing partners.
Reports suggest that NVIDIA is already exchanging SOCAMM prototypes and conducting rigorous performance tests with its partners, which also include some domestic firms like Simmtech and TLB. The current goal is to start mass production before the end of this year. If successful, this memory innovation will not only enhance AI computing performance but also reshape the future of high-performance computing and personal AI systems.
