Nvidia was working to diversify its sourcing of HBM3 and 2.5D packages used in AI GPUs for data centers, TheElec has learned.
The US chip giant was negotiating deals with potential suppliers, which includes Samsung, sources said.
Currently, Nvidia, for its A100, H100, and other AI GPUs, uses TSMC for the front-end process of manufacturing the wafers and 2.5 package work.
The HBM, or high bandwidth memory, chips used with the AI GPUs are being exclusively supplied by SK Hynix.
However, TSMC doesn’t have the capacity to handle all the workload of the 2.5D packaging these chips need to go through.
Nvidia is holding talks on volume and price with secondary and alternative suppliers, sources said.
These include Amkor Technology of the US and SPIL of Taiwan. SK Hynix will continue to supply the HMB3 used.
Samsung’s advanced package team (AVP) is another potential supplier. This team was formed late last year by the South Korean tech giant to expand revenue from chip packaging.
AVP team can receive the AI GPU wafers Nvidia purchases from TSMC; it can then procure HBM3 from Samsung’s memory chip business unit and use its own I-Cube 2.5D packaging to finish the work.
Samsung made this proposition to Nvidia, while also adding that it can commit a substantial number of engineers to the project. Samsung also offers to design the interpose wafer for Nvidia as well.
If the South Korean tech giant is expected to handle around 10% of Nvidia’s AI GPU package volume through the deal, sources said.
However, Samsung will have to meet Nvidia’s requirements and pass the quality test for its HBM3 and 2.5D package, they added.
Samsung is planning to start production of HBM3 later this year. Earlier this month, Samsung CEO and chip boss Kyung Kye-hyun, through the company’s internal chat messenger, shared that the company’s HBM3 was evaluated as excellent by a customer. Kyung said that he expected HBM3 and HBM3P to contribute to the increase in profits of the chip division starting next year.
Meanwhile, TSMC was also planning to expand its 2.5D packaging capacity by over 40% to meet the rising demand from Nvidia.
This means unless Samsung passes Nvidia’s evaluation quickly, it might fail to clinch the deal.
Samsung is also working to develop a bump-less package by 2025. The package is aimed at high-layer HBM and can help lower the height of the packages. The technology is also called Cu to Cu, direct bonding or hybrid bonding.