The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

Samsung Unveils SOCAMM2 Memory for Agentic AI Servers

By The Daily Byte ·21 Nov 2026 ·4 min read
Samsung workers in cleanroom suits inspecting green memory modules in a semiconductor fabrication plant

Samsung Electronics has introduced SOCAMM2, a new server memory module designed for the era of agentic AI, in which artificial intelligence systems set goals, reason through problems and act on their own over long periods of time.

The company presented the module in a post on its semiconductor technology blog on 19 November, describing it as a memory solution built around low-power DRAM, or LPDDR, rather than the conventional DDR memory used in most servers today. SOCAMM2 stands for Small Outline Compression Attached Memory Module.

Agentic AI places very different demands on memory than earlier workloads. As Samsung put it, longer context windows, repeated rounds of reasoning and many tasks running at the same time are making memory performance and power efficiency increasingly important. The company identified four essential requirements for AI server memory in this new era: high capacity, high bandwidth, power efficiency and space efficiency.

A detachable design instead of soldered memory

The key difference from existing LPDDR server memory is that SOCAMM2 is a detachable module. Until now, LPDDR chips used in servers have typically been soldered directly onto the motherboard, which means the memory cannot be replaced or upgraded without replacing the whole board.

SOCAMM2 keeps LPDDR’s high bandwidth and low power consumption but packages it as a removable module, so memory can be swapped or upgraded independently of the motherboard. That simplifies servicing and extends the useful life of server hardware, which matters when data centre operators are refreshing large fleets of machines to keep up with AI demand.

The module also uses a horizontal form factor, lying flat rather than standing upright like a conventional RDIMM. Samsung said the layout makes better use of the space around processors and accelerators, and allows heatsinks to be placed so that air can flow from the front to the back of densely packed AI servers.

153.6 GB/s and a CES award

Samsung said the LPDDR5X version of SOCAMM2 delivers up to 153.6 GB/s of bandwidth and more than 70 percent better energy efficiency than conventional DDR5 RDIMMs. Those are the company’s own comparison figures against its cited DDR5 reference, not independent measurements of complete AI servers.

The LPDDR5X SOCAMM2 has been named a CES 2027 Innovation Awards honouree, alongside two other Samsung semiconductor products: a 512GB DDR5 registered memory module and UFS 5.0 mobile storage. The Consumer Technology Association’s awards recognise product design and engineering ahead of the CES trade show in Las Vegas in January.

Samsung said customer samples of SOCAMM2 are already being supplied, and that it is working with Nvidia to optimise the memory for the chipmaker’s next-generation Vera Rubin AI inference platform. Looking further ahead, the company plans a version based on the next-generation LPDDR6 standard, which it said would push bandwidth, capacity and power efficiency higher still.

Why memory has become an AI battleground

Memory has become one of the most contested parts of the AI hardware supply chain. Training and running large AI models needs vast amounts of fast memory close to the processors, and the power consumed by that memory is now a major cost and cooling problem for data centre operators.

High-bandwidth memory, or HBM, has dominated AI accelerators, but it is expensive and in short supply. LPDDR-based modules such as SOCAMM2 offer an alternative for the CPU servers and inference systems that surround the accelerators, where capacity and power efficiency matter as much as raw speed.

JEDEC, the industry standards body, is also nearing completion of a common standard for SOCAMM2 modules, which should make it easier for server makers to adopt the form factor across different suppliers. Samsung’s move, combined with rival Micron already sampling its own 192GB SOCAMM2 modules, suggests the format is moving from a niche idea to a mainstream part of AI data centre design.