DRAM

When memory changes the equation: What we learned putting LPDDR to work in the data center

Khayam Anjam

Neon-lit data center server corridor

As AI workloads increasingly dominate data center requests, the constraints that matter most for general purpose and AI servers are shifting. Power and thermal budgets, not just raw compute, increasingly decide how much useful work a rack can do. That’s exactly the pressure point where LPDDR5X and the new SOCAMM2 modular form factor start to provide interesting data center advantages. SOCAMM2 brings the efficiency and density of low power DRAM into a modular, serviceable module designed for servers, which opens the door to system architectures that simply weren’t practical before.

But a good idea on a slide isn’t the same as a proven one. So we set out to answer a harder question: when you deploy this memory against real data center workloads, what difference does it make?

To do that credibly required close collaboration: Micron’s Data Center Workload Engineering team worked shoulder-to-shoulder with engineers at Meta, using Meta’s open-source DCPerf benchmark suite, which reflects production environments across a hyperscale fleet. That partnership is what makes the results worth paying attention to: these measurements are grounded in workloads that reflect real-world data center environments.

We looked at the dimensions that decide whether a memory technology belongs in the data center: bandwidth, latency, capacity, and power efficiency. Our work characterized how LPDDR5X, including high-capacity configurations, could affect each. Some of what we found surprised us, particularly what happens when you give capacity-bound workloads enough memory to operate.

The takeaway is simple: the memory conversation in the data center is no longer just about the fastest or the biggest. It’s about matching the right memory to the right workload, and SOCAMM2 gives system designers a genuinely new option to work with.

I won’t spoil the details here. The full story (methodology, workloads, and numbers) is in the whitepaper we published with Meta. If you’re wrestling with power ceilings, capacity limits, or the economics of scaling AI infrastructure, I think it’s worth your time.

Systems Performance Engineer, Micron

Khayam Anjam

Khayam Anjam is a Staff Engineer at Micron Technology. As part of Micron’s Data Center Workload Engineering team, he works on characterizing and optimizing modern AI workloads across GPU-accelerated and data center platforms. He also contributes to technical papers, internal reports, and customer-facing collateral. Prior to Micron, Khayam designed AI systems for computer vision and anomaly detection. At Dell Technologies, he authored six patents while contributing to ML, analytics, and reliability initiatives. He brings more than 10 years of industry experience spanning performance engineering, generative AI, and distributed systems. Khayam holds an M.S. in Computer Engineering from Clemson University.

Related Blogs