The new part, based on LPDDR5X technology, promises 50 per cent more capacity than its predecessor without bulking up the footprint.
Micron reckons this upgrade chops time to first token by more than 80 per cent in real-time inference workloads, thanks to its 1-gamma DRAM process. The company claims it improves power efficiency by more than 20 per cent, a figure that sounds modest until you multiply it across racks stuffed with 40 terabytes of CPU-attached memory.
SOCAMM2’s modular design also makes life easier for engineers sweating it out on the server floor. It is supposedly simpler to service and leaves room for future capacity bumps.
The launch comes off the back of a five-year tie-up with Nvidia, which helped Micron get its low-power DRAM tech off the ground. SOCAMM2 takes LPDDR5X’s smartphone-grade thriftiness and repackages it for AI servers that need to sling data at terrifying speeds without melting the power grid.
Micron Cloud Memory Business Unit senior vice president and general manager Raj Narasimhan said: “As AI workloads become more complex and demanding, data centre servers must achieve increased efficiency, delivering more tokens for every watt of power.”
“Micron’s proven leadership in low-power DRAM ensures our SOCAMM2 modules provide the data throughput, energy efficiency, capacity and data centre-class quality essential to powering the next generation of AI data centre servers.”
The company insists the modules are ready for serious data centre use, claiming they meet the reliability and quality standards expected in hyperscale environments. Each module, a third the size of a standard RDIMM, sips two-thirds less power while pushing comparable bandwidth. That makes it a solid choice for liquid-cooled servers where every watt and cubic centimetre counts.
Micron has been busy in the standards arena, helping JEDEC nail down the SOCAMM2 spec and pushing for wider industry adoption of low-power server memory. Samples of the 192GB SOCAMM2 are already in customers’ hands, clocking speeds up to 9.6Gbps, with mass production expected to align with upcoming AI platform launches.