🧠 Food for thought

Implications, context, and why it matters.

Vera Rubin’s memory needs help explain Nanya’s entry

• The LPDDR5X supply will support Nvidia’s next-generation Vera Rubin platform, not Blackwell 1.
• Each Vera Rubin Superchip pairs a CPU with a GPU and carries 1.5 terabytes of memory. That is three times the capacity of the prior system 2.
• The need remains even with 90% key-value (KV) cache compression, which cuts memory used for stored context. AI systems still need more hardware for agentic AI workloads 2.
• Nvidia is adding suppliers such as Nanya to help prevent future shortages for Vera CPUs 2.

Nanya’s deal marks a change in server memory design

• Nanya’s addition to the supply chain tracks the move to SOCAMM2, a newer server memory module format.
• SOCAMM2 puts low-power LPDDR5X memory into a modular, replaceable package instead of soldering it to system boards. That makes it better suited for data centers 3.
• SOCAMM2 delivers more than twice the bandwidth and over 75% better energy efficiency than registered dual in-line memory modules (RDIMMs) 4.
• JEDEC, the semiconductor standards body, is standardizing this work through its JESD328 CAMM2 specification. The change moves server memory away from older designs to meet AI’s power and performance demands 3.