AI Architecture Shifts Drive Flash Demand as Macronix Expands Tiered Memory Strategy
Frontier AI models
Expanding AI servers and inference models are reshaping memory requirements. As HBM capacity pressures mature processes, high-density NOR Flash, HBF, and tiered storage architectures emerge as key strategies to ease data transfer bottlenecks.
As AI model scale and inference workloads surge, major semiconductor manufacturers are shifting capacity toward High Bandwidth Memory (HBM), squeezing mature processes. According to Macronix Marketing Director Lin Min-cheng, modern AI servers now require over 1,000 NOR Flash units per rack compared to 600 to 800 in previous generations. To tackle expanding KV caches and data transfer bottlenecks, the industry is exploring tiered storage architectures, such as 3D NAND-based HBF and dual-tier SSDs, alongside memory-centric computing.
Key takeaways
- Rising hardware complexity has driven NOR Flash usage in the latest AI server racks to over 1,000 units.
- To address growing KV caches and cost constraints, the industry is exploring HBF and dual-tier SSD architectures to offload HBM data.
- Capacity displacement has tightened mature Flash supply, prompting Macronix to expand its high-density NOR Flash and 3D NAND footprint.
Source：EE Times Taiwan 電子工程專輯 Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.