Micron Unveils Nvidia-Friendly SOCAMM

On the heels of announcing its 1-gamma DRAM that put a lot of emphasis on reduced power consumption, Micron Technology unveiled its 128GB small outline compression attached memory module (SOCAMM).

Leveraging LPDDR5X memory and co-developed with Nvidia, Micron’s SOCAMM will be in the upcoming GB300 Blackwell Ultra Superchip. In a briefing with EE Times, Praveen Vaidyanathan, Micron’s VP of DRAM process integration, said the value of LPDDR—which is a mainstay in mobile phones—is its lower power. Micron has been working on LPDDR technologies aimed toward the data center for four years.

“SOCAMM is essentially a modular version,” he said. “It is the highest capacity modular memory that we have built for data centers.”

Micron did add some features that make it different from the LPDDR that shows up in consumer products, including the test flows, but ultimately the goal is to leverage lower power in the data center as memory, performance and capacity are scaled up to meet AI workload demands, Vaidyanathan said.  A SOCAMM consumes one-third the power compared to standard DDR5 registered dual in-line memory module. It is also easier to stack together, he added.

“You can do this through conventional wire bond technology.”

A big difference between consumer applications like smartphones that use LPDDR is that they are more tolerant than a data center environment that requires high reliability, Vaidyanathan said.

“There’s more wiggle room in a consumer smartphone when you’re doing an LPDDR versus the kind of workloads in a data center.”

The server industry is used to a modular memory solution, Vaidyanathan added. “One of the key requirements of this design was we wanted this to occupy less space.”

Newer platforms have very high memory density in terms of gigabytes per millimeter square, so SOCAMM was designed to allow for 32GB in a package. “You put four of them on a PCB to give you 128 gigabytes,” he said. “You pack a lot of capacity in much smaller space. It gives you a much higher performance for a given module.”

Vaidyanathan noted that PCB design and channel design has evolved enough where LPDDR can function in a modular environment to handle the loading that comes from channels, but the SOCAMM did require a package, module and server channel design understanding.

“That was a huge part of enabling an LPDDR that can function at the bandwidth requirements of these high-performance systems.”

Micron’s introduction of a SOCAMM that works with Nvidia’s GB300 Blackwell Ultra Superchip comes at a time when HBM is in high demand, according to recent research released by IDTechEx.

“GPUs demand very high bandwidth, which is delivered from the HBM,” Vaidyanathan said. “But for an AI accelerator you also need very high capacity, which we think the low power DRAM delivers.”

He added that Micron sees LPDDR as a partner solution for AI infrastructure that supports the movement of large quantities of data, along with HBM, DDR and GDDR, depending on performance, capacity and power needs.

Micron claims to be first out of the gate to deliver memory that supports Nvidia’s upcoming superchip, but Jim Handy, principal analyst with Objective Analysis, told EE Times in an interview that it will not be the only vendor that steps up the plate.

“Nvidia’s orchestrating an effort with everybody. It’s just that Micron’s ahead of everybody else,” he said. “That’s something that Micron has started to take very seriously—putting itself in the position of being a leader.”

The adoption of LPDDR in the data center mirrors that of SSDs, Handy added, in that power consumption concerns were already rearing their head as early as the late ’90s, and that by adding SSDs into a system, power hungry DRAM could be reduced without compromising performance.

He said the LPDDR specification was designed with the idea that it would eliminate the hassles that come with DIMMs and that it would be soldered down next to the processor. Therefore, it was relegated to going into small systems like cell phones. Handy said a SOCAMM enables LPDDR to be expandable.

Micron’s SOCAMM is an example of how the technology is not just for PCs, but also something that can help in the data center. “Nvidia is really championing this,” Handy said.

Micron recently announced it is shipping samples of its 1γ (1-gamma), sixth-generation (10-nm class) DRAM node-based DDR5 memory. The 16Gb DDR5 memory is designed to offer speed capabilities of up to 9200MT/s, a 15% speed increase compared to its predecessor, while also reducing power consumption by 20% by using next-generation high-K metal gate CMOS technology paired with design optimizations.