A third-generation SOCAMM module is reportedly being prepared for 2028 systems, signalling that Nvidia’s compact memory format could expand beyond its original AI-server focus.
SOCAMM is moving from an experimental AI-server memory concept to something that now appears to have a multi-generation roadmap. According to ComputerBase, the third version of the standard is already being prepared for systems expected from 2028 onwards, with Nvidia’s Rosa family and the Rosa-Feynman platform seen as the most immediate targets. AMD, meanwhile, also appears to be aligning future server designs with the same direction. Tom’s Hardware has previously reported that Nvidia has been working with SK hynix, Micron and Samsung on SOCAMM as a compact, high-performance module intended for AI workloads, with prototypes already circulating for testing.
The format itself is a deliberate break from conventional DIMMs. SOCAMM stands for Small Outline Compression Attached Memory Module, and it uses low-power DRAM rather than the taller stick modules seen in standard desktop and server systems. ComputerBase notes that the currently specified module measures 90mm by 14mm and is mounted flat against the board rather than inserted vertically. The design uses four memory chips on one side, while the reverse carries 694 I/O contacts. It is secured with three screws to maintain even pressure across the length of the module.
That physical layout reflects a broader design goal: more capacity in less space, while preserving efficiency. Tom’s Hardware has reported that the first commercial SOCAMM modules use LPDDR5X and are aimed at AI and low-power servers, with capacities reaching 128GB in some configurations. The appeal is clear in data-centre systems, where power draw, board space and serviceability all matter. The detachable format is meant to make upgrades easier than soldered memory, while still delivering a smaller footprint than traditional server memory arrangements.
The pace of development has been unusually quick. ComputerBase points out that the first SOCAMM generation, announced in spring 2025, never progressed far before the industry shifted attention to SOCAMM2 later that year. Until recently, Nvidia was effectively the only major customer. AMD, by contrast, had not initially planned around the standard, but is now expected to support it with Verano, its server CPU due from early 2027. That shift suggests the format is no longer a niche Nvidia-specific project, but a wider platform choice for future AI servers.
The next step is already taking shape. According to South Korean reports cited by ComputerBase, PCB supplier TLB is working on SOCAMM3 and is in discussions with two customers, although final specifications have not been settled. Mass production is nevertheless being targeted for 2028. TLB is said to be facing manufacturing challenges because the module’s stacking process is more complex than before, requiring additional layers and tighter control over the substrate. The company’s role also highlights an economic reality: SOCAMM modules are more complex, and therefore more valuable to produce, than standard UDIMMs or RDIMMs.
LPDDR6 is widely expected to be the likely companion memory for SOCAMM3, given the standard’s emphasis on low power and compact packaging. That would fit the trajectory already visible in the roadmaps of Nvidia and AMD. TechYorker reported that Nvidia’s 2028 Feynman architecture is paired with the Rosa CPU platform, while AMD has already confirmed that its next-generation Florence EPYC family, based on Zen 7, will include what it calls next-generation memory support from late 2028. Taken together, the signals point to SOCAMM becoming a long-term option for high-density AI infrastructure rather than a transitional experiment.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
