SK hynix is reportedly strongly considering using Taiwan Semiconductor Manufacturing Co.'s (TSMC) 3-nanometer (nm; one-billionth of a meter) process as the main option for the "logic die," which serves as the brain of its seventh-generation high-bandwidth memory (HBM) HBM4E. The move appears aimed at gaining a performance edge by applying advanced processes not only to the DRAM stacked in HBM but also to the logic die that handles HBM's computational processing. The HBM4 (sixth-generation HBM) that SK hynix will mass-produce this year is said to lag Samsung Electronics in certain performance metrics.
According to industry sources on the 20th, SK hynix plans to use a 10-nanometer-class sixth-generation (1c) DRAM process for the "core die," the DRAM stacked in HBM4E, and apply TSMC's 3 nm process to the logic die. The HBM4 that SK hynix is supplying to Nvidia this year uses a core die built on a 10-nanometer-class fifth-generation (1b) DRAM process and a logic die made with TSMC's 12 nm process. Samsung Electronics has equipped HBM4 with a core die built on a 10-nanometer-class sixth-generation (1c) DRAM process and a logic die using a 4 nm process.
SK hynix's strategy is to significantly boost performance in next-generation HBM by applying finer processes. In semiconductors, processes become more miniaturized with each generation. Applying fine processes narrows the linewidth of semiconductor circuits, shortening the distance electrons travel to increase operating speed, while lowering the voltage required for operation to improve power efficiency.
While SK hynix is said to have solidified its No. 1 position in the HBM4 market by supplying the largest volumes to Nvidia, it is reportedly evaluated as trailing Samsung Electronics in performance. Samsung Electronics has said it applied a more advanced process than SK hynix to HBM4, demonstrating a performance edge and achieving the industry's first mass-production shipment.
With HBM4E, SK hynix appears to have played a bold card by using a 3 nm process for the logic die to overturn such evaluations. If HBM4 focused on stability by using proven processes, HBM4E is positioned to dominate even in performance and secure an advantage in the technology race.
Starting with HBM4E, the "custom HBM" market—where the logic die is designed and delivered in the circuit configuration desired by the customer—is expected to fully take off, raising the possibility that various foundry (semiconductor contract manufacturing) processes will be applied to logic dies. Still, SK hynix appears to be pursuing a plan to use the 3 nm process as its main option. For major customers with large supply volumes, such as Nvidia, the company plans to provide the highest-performing products. HBM4E will be used in the top-tier version of Nvidia's next-generation AI chip "Vera Rubin," the Vera Rubin Ultra.
A semiconductor industry official said, "For custom HBM4E, because the logic die is manufactured in the form the customer wants, options such as 3 nm and 12 nm are being considered," adding, "However, for the logic die in HBM4E that will be supplied in the largest share to the market, the 3 nm process will likely be the main choice."
Meanwhile, not only Nvidia but also AMD and Google have said they will equip their next-generation AI chips with HBM4E, suggesting competition surrounding HBM4E will intensify further. A semiconductor industry official said, "Samsung Electronics signaled its intent to preempt the next-generation market by unveiling HBM4E at Nvidia's annual flagship developer conference, GTC 2026," adding, "SK hynix appears to be pursuing a strategy to overwhelm even in performance with next-generation HBM."