Industry news--Karmin Electronic (HK) Limited

Samsung’s hybrid bonded HBM will be extended to 2029 and will debut alongside NVIDIA’s Feynman platform
2026-07-23

According to foreign media Wccftech, as high bandwidth memory (HBM) continues to evolve towards higher stack layers and higher bandwidth, the market generally believes that hybrid bonding is the key packaging technology for the next generation of HBM. However, according to industry reports in South Korea, Samsung may postpone the mass production of hybrid bonded HBM to 2029-2030 and synchronize it with Nvidia's next-generation Feynman AI GPU platform.

According to the data, the introduction of hybrid bonding technology will remove the micro bumps between the DRAM chips in the HBM stack and switch to direct copper to copper (Cu to Cu) bonding, which can significantly shorten the distance between chip layers, further improve bandwidth, reduce power consumption, and improve heat dissipation. Therefore, it is considered a key technology.

According to sources, Samsung's evaluation of hybrid bonding as truly mature for mass production will take place between 2029 and 2030, and is expected to be launched in conjunction with Nvidia Feynman AI GPU.

As the next generation AI platform for NVIDIA Rubin Ultra, it is rumored that the Feynman platform will adopt a 3D stacked architecture, with Intel responsible for some advanced packaging and equipped with customized high bandwidth memory (Custom HBM). Market speculation suggests that Feynman may adopt HBM5 or a customized version developed based on HBM5.

It is reported that Samsung will start installing hybrid bonding equipment at its Pingze factory as early as the end of this year, but initially it will mainly be used for technical verification and production line construction. The actual mass production still needs to wait until 2029 to 2030.

However, previous reports have indicated that hybrid bonding can be introduced as quickly as HBM4E. It is reported that Samsung is currently purchasing about 50 hybrid bonding devices from Dutch semiconductor equipment manufacturer BESI. But Samsung hopes that the devices can be customized, and of course, it does not rule out the possibility of evaluating the procurement of local equipment factories in South Korea.

As for customized HBM, it is based on the 3D System in Package (SiP) architecture, replacing the traditional HBM's Base Die with the customer's own Logic Die, making the integration of memory and computing chips more closely, further improving AI computing efficiency and system performance.

It is worth noting that there have been recent reports that JEDEC is considering relaxing the thickness specifications for HBM packaging, allowing HBM4E to still use existing hot press bonding technology and reducing the urgency for industry players to introduce hybrid bonding in the short term. Major companies such as Samsung and SK Hynix may postpone the full commercialization of hybrid bonding to the HBM5 generation to coincide with the release of Nvidia Feynman AI GPU.

Prev：Domestic large model Kimi K3 can independently design AI chip: 45nm process with 100MHz frequency Next：There is no next post! !