Samsung Ramps Up HBM4 Chip Production for NVIDIA
Samsung is reportedly accelerating its production of HBM4 memory chips, building out its hybrid bonding manufacturing lines earlier than planned at the request of NVIDIA. This move is part of Samsung's strategy to achieve supremacy in the high-bandwidth memory market. The decision reflects intense competition and high demand for advanced memory used in AI accelerators and high-performance computing.
- The JEDEC standards body finalized the HBM4 specification (JESD270-4), which doubles the memory interface width from 1024 bits in HBM3 to 2048 bits. This architectural change allows HBM4 to achieve a bandwidth of up to 2 TB/s per stack, a significant increase over the approximately 1.2 TB/s of HBM3E. - Samsung's adoption of hybrid bonding is a key technological shift for HBM4 manufacturing. This technique directly bonds copper-to-copper, eliminating the need for microbumps, which can improve thermal performance and signal integrity, though it is a more expensive process. - Competitor SK Hynix, the current HBM market leader, is developing an advanced version of its existing Molded Reflow Molded Underfill (MR-MUF) process for HBM4, keeping hybrid bonding as a backup plan to avoid the high capital investment associated with the new equipment. - The demand for HBM4 is largely driven by NVIDIA's next-generation "Rubin" AI accelerator platform, which is expected to launch in 2025. This has prompted memory manufacturers like SK Hynix and Samsung to accelerate their HBM4 mass production timelines. - NVIDIA is expected to source HBM4 from all three major suppliers—Samsung, SK Hynix, and Micron—as no single company can meet the total demand for the Rubin platform. TrendForce analysis suggests Samsung may be the first to secure certification due to its product stability. - To manage the supply, NVIDIA is reportedly considering a "dual-bin" strategy, using top-tier HBM4 chips with speeds up to 11.7 Gbps in premium systems and lower-tier versions around 10 Gbps in other models. Samsung's HBM4, which recently entered mass production, is positioned for the top performance tier. - Following the initial HBM4 launch, an enhanced "HBM4E" version is already on roadmaps, with Samsung expecting to sample it in the second half of 2026 and Micron targeting a late 2027 release. There is also a trend toward co-developing custom HBM products for major clients like NVIDIA, which could shift production capacity away from the standard memory market.
