Samsung to double HBM4 and HBM4E DRAM output for AI chips
Expansion targets 12-layer+ stacks as demand for high-bandwidth memory surges.
1Output doubling strategy
Samsung plans to more than double production of its HBM4 and HBM4E high-bandwidth memory chips in 2025. The expansion focuses on 12-layer and higher stack configurations, crucial for AI accelerators. This follows the company's February launch of 10nm-class sixth-generation HBM4 and May's 12-layer HBM4E samples provided to clients like Nvidia. The move positions Samsung to capture growing AI chip market demand.
2Glass carrier criticality
HBM production requires specialized glass carriers to prevent wafer bending during thinning processes. These temporary supports become more vital as stack counts increase. Samsung will boost outsourced glass carrier cleaning from 20,000 monthly sheets in 2024 to 50,000 in 2025 - a 2.5x increase from this year's already doubled volume. The carrier scaling directly enables higher HBM output.
3Manufacturing technology demands
HBM's stacked DRAM architecture necessitates advanced wafer thinning and warpage control technologies. As layers multiply to 12+, process precision becomes paramount. Samsung's 4nm base die technology and 10nm-class 1c DRAM form the foundation for these complex multi-layer structures. The company's ability to master these processes determines its HBM4 family yield rates.
4Production volume projections
Industry analysts forecast Samsung's total HBM production will grow nearly 40% to ~250,000 wafers annually by 2025. The HBM4 family's share will jump from 40% to 80% of output as HBM4E ramps up. This product mix shift toward premium HBM4 variants reflects Samsung's strategy to prioritize high-value segments in the competitive AI memory market.
5Market context
Samsung's expansion comes amid intense competition in AI-optimized memory. The HBM market has become strategic as generative AI workloads demand unprecedented memory bandwidth. Samsung's HBM4 push aims to challenge SK Hynix's early lead in AI memory solutions. The production doubling signals confidence in sustained AI infrastructure investment cycles.
6Supply chain implications
The HBM output increase requires coordinated scaling across Samsung's specialty materials ecosystem. Glass carrier suppliers face particular pressure to meet 2.5x higher demand while maintaining precision standards. This vertical integration challenge highlights how advanced memory production increasingly depends on materials science breakthroughs beyond semiconductor fabs.
7Technology roadmap
Samsung's HBM4 family represents its sixth (HBM4) and seventh-generation (HBM4E) high-bandwidth memory. The 12-layer HBM4E samples delivered to Nvidia demonstrate progress toward mass production. Each generation increases bandwidth and power efficiency - critical metrics for AI training clusters where memory can bottleneck system performance.
8Strategic importance
Samsung's HBM4 bet reflects memory's central role in AI hardware ecosystems. As AI models grow exponentially, memory bandwidth becomes the critical path for accelerator performance. Controlling advanced HBM production allows Samsung to influence AI infrastructure development timelines and capture premium pricing in this high-growth specialty memory segment.
Original article: Hacker News — Samsung to double HBM4 and HBM4E DRAM output for AI chips