Samsung's Hybrid Bonding technology is expected to be used in NVIDIA's next-generation AI accelerator, Feynman, with the chip featuring custom HBM.

BlockBeats News, July 21st, Samsung Electronics has started construction of the Die-to-Wafer (D2W) hybrid bonding mass production line at its Pyeongtaek P5 plant, aiming directly at the advanced packaging of the next-generation HBM and logic semiconductor. The core equipment supplier is the Netherlands-based Besi. Samsung plans to purchase about 50 hybrid bonding machines, with each machine priced at around 60 billion Korean Won (approximately $43 million), which is twice the price of competing products. However, due to Samsung's request for customized modifications to the equipment architecture, the negotiation progress has been slow. Besi also supplies equipment to TSMC and Micron at the same time, so they are cautious about customizing the machines solely for Samsung.

Samsung is also considering domestic alternatives: its subsidiary SEMES has completed the development of D2W hybrid bonding machines and sent samples for verification earlier this year; Hanwha Semitech completed the development of the second-generation 'SHB2 Nano' in February, and in April, samples were sent to SK hynix. Their cluster system, integrating EFEM, plasma activation, and cleaning modules, has formed a differentiating competitive advantage.

This technology uses hybrid copper bonding to replace traditional thermocompression bonding, allowing direct bonding of copper to the dielectric material, significantly shortening the interconnection length. The core application scenario is for custom HBM, where the HBM DRAM layer is directly vertically stacked on a logic die containing customer computing circuits and IP to form a 3D system-level package.

Citrini analyst Jukan stated that hybrid bonding is expected to be used in NVIDIA's next-generation AI accelerator Feynman, which will adopt custom HBM. The timeline for technology implementation has been "pushed back from HBM4E." Internally at Samsung, it is estimated that large-scale application may be achieved around 2029 to 2030, aligning with the expected window for NVIDIA's Feynman.