Micron and SanDisk Propose New Memory Architectures to Challenge HBM
nashnova research
Micron and Sandisk each unveiled a new memory architecture attacking HBM from different angles — bandwidth and capacity — signaling that AI inference demand is outgrowing what a single memory standard can cover.
What is HBM, and why is it being challenged?
HBM — high bandwidth memory, a packaging method that vertically stacks multiple DRAM dies and bonds them close to the processor — is the default memory in today's AI chips; every Nvidia GPU ships with it.
But AI inference workloads are stretching both data volume and speed at once. HBM is fast, yet capacity is limited; stacking more layers drives up cost and power.
This means → a single HBM roadmap is increasingly unable to deliver "fast, large, and cheap" simultaneously. The market needs alternative or complementary paths.
What problem does Micron's approach solve?
Micron is developing "tightly coupled DRAM" — an architecture that places DRAM directly on the processor, shortening the data-travel distance — targeting bandwidth density.
Headline specs: theoretical bandwidth exceeding HBM by 10×, with significantly lower energy per bit.
In plain terms = Micron is not trying to store more; it wants to move data faster and burn less power doing it. The bottleneck it attacks is the data channel between processor and memory.
What problem does Sandisk's approach solve?
Sandisk introduced High Bandwidth Flash (HBF) — a NAND-flash-based package that pulls flash storage closer to the processor — targeting capacity-to-cost ratio.
Key numbers: bandwidth claimed to match HBM, while offering 8–16× the capacity at a similar price.
This means → Sandisk is aiming at a different pain point. Inference models keep growing, and the storage they need far exceeds HBM's ceiling. Flash's natural capacity advantage fills that gap.
Do both paths point to the same signal?
Micron attacks bandwidth density; Sandisk attacks capacity-to-cost. Completely different technologies, yet both rest on the same assessment: AI inference storage demand has outgrown what HBM alone can cover.
This reflects a shift in AI-memory competition — from "who has the most HBM capacity" to "what exists beyond HBM." The race itself is forking.
Neither company has disclosed a mass-production timeline or commercialization details. Whether yields and costs can deliver on theoretical specs remains the key validation step.
市场有风险，内容仅供研究参考，不构成投资建议。