AMD and Qualcomm have spotted Nvidia’s SOCAMM memory and decided it looks worth nicking for their own AI gear.
A Hankyung report says Nvidia, Qualcomm, and AMD are exploring SOCAMM modules for upcoming AI racks, as agentic AI apps are making memory the limiting factor.
SOCAMM started life as a memory standard supposedly tailored for Nvidia, with Team Green first in line to adopt it.
It is based on LPDDR DRAM, the type of memory you typically see in mobile and low-power devices, but SOCAMM is upgradable rather than soldered to the PCB.
That makes it a handy sidekick for HBM, allowing you to shift memory-bound tasks without requiring you to replace an entire board when you need more capacity.
AMD and Qualcomm are considering a different layout from Nvidia, using a “square” module with two DRAMs split across two rows.
The idea is to move power control via the PMIC directly onto the module, which should improve regulation at extreme speeds and reduce motherboard complexity by eliminating redundant power circuitry.
If SOCAMM spreads beyond Nvidia, DRAM usage for the format will rise as well, because agentic AI requires fast, short-term storage alongside HBM.
SOCAMM can push terabytes of memory per CPU, keeping millions of tokens “active” in working memory, even if its throughput is slower than HBM.
Nvidia is expected to ship SOCAMM 2 with Vera Rubin AI clusters, and if AMD and Qualcomm follow through, the memory arms race gets even messier.
