SOCAMM2
As the computing power for AI large model training and inference surges toward GW (gigawatt)-scale data centers, the bottleneck for GPU computing output is no longer just the physical computing limits of the chips themselves, but the rapidly escalating power consumption and cooling costs in data centers.
Recently, tech giants such as NVIDIA, Microsoft, AMD, and Qualcomm have successively secured orders and capacity for a new type of compression-attached memory module—SOCAMM2 (Small Outline Compression Attached Memory Module 2).
This low-power memory technology, originally derived from smartphones and ultrabooks, is tearing open a gap in the server storage market in an unexpected way.
01. SOCAMM2 Enters AI Data Centers
In the traditional physical architecture of data centers, DDR5 RDIMM holds an absolute dominant position in server main memory, thanks to its easy plug-and-play capability, high capacity, and mature supply chain ecosystem. However, in inference scenarios centered on large models with hundreds of billions of parameters, this long-standing rule set is failing.
Given the proliferation of long context windows, the memory capacity occupied by KV Cache (Key-Value Cache) is growing exponentially. Relying entirely on expensive and extremely capacity-constrained HBM would make commercialization costs unbearable; conversely, handing it all over to traditional DDR5 RDIMM would cause its high power consumption and transmission latency to quickly exceed the power budget of the rack.
The introduction of SOCAMM2 is essentially a technological graft that "modularizes high-energy-efficiency mobile memory for server applications."
To understand this transition more intuitively, we can make a popular analogy:
Traditional DDR5 RDIMM is like a heavy-duty diesel locomotive; although its carriages can be detached and replaced at any time (high scalability), its fuel consumption is astonishing during operation (extremely high power consumption);
On-board LPDDR5X in mobile phones is like a supercar engine, offering excellent performance and extreme fuel efficiency, but it is directly welded onto the chassis. If a single spark plug fails, the entire car must be scrapped (non-maintainable);
SOCAMM2, on the other hand, equips this supercar engine with "magnetic quick-release buckles." It retains the ultra-low power consumption and high transmission rate of LPDDR5X, while utilizing screw-locked compression attachment technology, allowing server O&M personnel to perform modular maintenance and upgrades just like replacing traditional memory sticks.
Based on the calculation data published by Micron and the specification requirements of JEDEC (the semiconductor standard organization), the comparison of core features between SOCAMM2 and traditional server memory is summarized as follows:
02. Global Supply Chain Synergy: From the Memory Manufacturers' Battle to the Connectors Taking the Baton
Judging from the actual moves in the global supply chain, the offensive and defensive battle surrounding SOCAMM2 has fully unfolded among memory manufacturers, chip designers, and component packaging houses.
1. Memory IDM Manufacturers: The Mass Production Race for 1γ/1c Advanced Nodes and High-Density Modules
Micron: In October 2025, it took the lead in sampling 192GB SOCAMM2 modules to customers; subsequently, in March 2026, it announced the launch of a 256GB SOCAMM2 module based on 32Gb single-die 1γ (1-gamma) node. Micron's empirical data on the Llama-3 70B model (500K context length) shows that offloading the KV Cache to 2TB of SOCAMM2 main memory can accelerate the Time to First Token (TTFT) by 2.3 times (reducing latency by over 80%), with operating power consumption being only one-third of that of RDIMMs with the same capacity.
SK Hynix: In April 2026, it announced the mass production of 192GB SOCAMM2 modules based on 1cm (6th generation 10-nanometer class) LPDDR5X, offering more than double the bandwidth of traditional RDIMMs and a 75% reduction in power consumption. SK Hynix officially pointed out that this product is specifically designed for the Vera CPU in NVIDIA's next-generation Vera Rubin platform.
Samsung: Simultaneously advancing the R&D of 10.7 Gbps high-speed LPDDR5X dies, and sampling server-grade CAMM2 and SOCAMM2 modules to cloud customers, actively competing for follow-up model orders from NVIDIA and Microsoft.
2. Computing and CPU Chip Designers: Comprehensive Adoption of the SOCAMM2 Memory Interface
NVIDIA: In its Vera CPU architecture specifically built for Agentic AI, it comprehensively introduces the SOCAMM2 memory subsystem (configuring 8 SOCAMM2 slots per CPU node). In June 2026, affected by the tight wafer capacity of LPDDR5X from memory manufacturers (with the manufacturers' supply fulfillment rate at about 60%), NVIDIA fine-tuned its procurement structure, reducing the packaging of some single SOCAMM2 modules from 4-layer stacking (192GB) to 2-layer stacking (96GB), to prioritize the shipment schedule of the entire Vera CPU machine under the constraint of limited dies.
AMD/Qualcomm/Microsoft: AMD has officially added support for the JEDEC JESD328 compression module specification to its EPYC roadmap; Qualcomm adopts the SOCAMM2 interface in its self-developed data center CPU platform; Microsoft Azure promotes the standardized procurement of SOCAMM2 from the perspectives of rack PUE energy saving and server height limitations.
Rambus: Launched a server chipset specifically built for SOCAMM2, covering Serial Presence Detect (SPD) Hub, Power Management IC (PMIC), and physical layer interface IP, taking the lead in the supporting host controller chip market.
3. Midstream Connectors, PCB, and OSAT Supply Chains: Deterministic Growth Brought by Physical Form Factor Changes
LGA Compression Connectors (Sockets): Traditional vertical DIMM slots are being eliminated, shifting towards 694-pin LGA compression sockets that extremely test contact resistance and flatness precision. Connector manufacturers such as Lotes, FIT, and Uconn have entered the relevant testing and mass production stages, with the value per unit of connectors significantly increased compared to traditional slots.
High-End PCB and Low-Dk Materials: High-speed transmission of 9.6 Gbps+ requires the module and motherboard PCB to have extremely low signal attenuation. PCB enterprises such as Unimicron and Compeq have increased the usage ratio of Low-Dk (low dielectric constant) materials to meet the high-density routing requirements of SOCAMM2 modules.
Advanced Packaging and Testing (OSAT): The dense stacking of multi-chip dies has enabled outsourced semiconductor assembly and test (OSAT) providers such as JCET, TFME, and ASE to gain deterministic order growth in the fields of server-grade LPDDR5X SiP (System in Package) and high-density module testing.
03. Mobile Phone "Granaries" Plundered by AI Servers
The explosion of SOCAMM2 in data centers is causing a certain degree of capacity squeeze on the consumer electronics supply chain.
In the past, over 85% of the core consumption channels for LPDDR5X were concentrated in high-end smartphones and ultrabook fields. However, as memory manufacturers prioritize the allocation of LPDDR5X wafer capacity from advanced nodes (such as 1β and 1γ) to the more profitable 256GB SOCAMM2 server modules, the supply-demand balance of mobile storage dies has been broken.
Industry analysis suggests that this shift may trigger a two-way chain reaction:
For the consumer electronics side: Smartphone manufacturers face dual pressures of rising quotes and prolonged lead times when procuring high-specification 16GB/24GB LPDDR5X dies. Some mobile phone brands are forced to delay memory specification upgrades in their mid-to-low-end product lines.
For the server side: Due to the extremely high yield requirements for high-stacking SOCAMM2, the effective shipment volume of high-capacity modules in the short term is still constrained by the manufacturers' packaging production lines. There are even instances in the industry where customers like NVIDIA, to ensure spot shipments, temporarily adjust their procurement structures and simultaneously purchase relatively affordable specification modules such as 96GB/128GB.
In summary, the rise of SOCAMM2 in data centers is not a simple "specification iteration," but the result of the explosive growth in AI computing power forcing a breakthrough in hardware architecture.
When traditional performance stacking hits the hard wall of physics and power consumption, combining mobile low-power technology with server modular design may become the key means for chip giants to break through the bottleneck.
In this storage revolution triggered by power consumption, whether it is upstream memory manufacturers, connector manufacturers, or downstream system integrators, they will all redefine their ecological niches in the AI industry chain.
#SOCAMM2 #Micron #NVIDIA