NVIDIA Adds Taiwan Memory Maker to Power Next‑Gen AI Chips
Taiwan’s Nanya Technology Joins Vera Rubin’s High-Performance Stack
NVIDIA is supercharging its Agentic AI platform, Vera Rubin, with a strategic expansion of its memory supply chain. The company has integrated LPDDR5X chips from Taiwan’s Nanya Technology, marking a first for its high-performance AI servers.
Memory Strategy: Power vs. Speed
Vera Rubin operates on a dual-memory architecture:
- CPUs leverage LPDDR5X for low-power efficiency and high density.
- GPUs rely on HBM4 for ultra-fast data transfer.
With HBM4 production restricted to a handful of suppliers, NVIDIA has diversified its LPDDR5X sources beyond Micron and SK Hynix, securing Nanya Technology as a new partner.
Exclusive Standards & Manufacturing Breakthroughs
Nanya’s chips meet Vera Rubin’s stringent AI performance requirements—a milestone many local Taiwanese manufacturers previously couldn’t achieve. TSMC’s advanced manufacturing support played a pivotal role, enabling Nanya to qualify for NVIDIA’s demanding platform.
Unmatched Performance Gains
The new LPDDR5X memory delivers dramatic improvements:
- Triple the memory capacity compared to older systems.
- 50% increase in bandwidth.
Superchip Specifications:
- 1.5 TB per chip
- 1.2 TB/s data transfer rate
Full Rack Capabilities:
- 400 TB total memory
- 315 TB/s aggregate bandwidth
Strategic Shift: AI’s New Focus on CPUs
As Agentic AI pivots from pure GPU acceleration to CPU-driven intelligence, securing multiple global memory suppliers is critical. NVIDIA’s decision to partner with a Taiwanese manufacturer underscores its commitment to reliable, high-performance memory diversification.
