Shulou(Shulou.com)11/24 Report--

CTOnews.com, July 20 (Xinhua)-- earlier this month, the Electronic Times said that Nvidia is considering outsourcing some of its AI GPU to Samsung Electronics due to the growing shortage of TSMC capacity.

According to TheElec, Nvidia is working to diversify the HBM3 and 2.5D packages used in its data center AI GPU. The chip giant is in talks with potential suppliers, including Samsung, according to sources.

According to CTOnews.com query, Nvidia A100, H100 and other AI GPU currently use TSMC's front-end processes for wafer fabrication and 2.5 packaging.

According to the report, the HBM memory used in Nvidia AI chips is provided exclusively by SK Hynix, but TSMC does not have the ability to handle all the workload of 2.5D packaging required for these chips.

According to sources, Nvidia is negotiating quantity and price with second-tier and alternative suppliers, including Amkor Technology (Andu Technology) and SPIL (Silicon Precision Industry). Of course, the new suppliers will not affect the status of SK Hynix.

Samsung's Advanced Packaging team (AVP) is said to be another potential supplier, which was set up at the end of last year to expand revenue from Samsung's chip packaging business.

The Samsung AVP team can receive AI GPU wafers purchased by Nvidia from TSMC, then purchase HBM3 from Samsung's memory chip business unit and use its own I-Cube 2.5D package for further work.

Samsung has made the proposal to Nvidia, according to sources, and Samsung says it can send a large number of engineers to participate in the project. In addition, Samsung also proposes to design interconnect chips for Nvidia.

If the deal goes through, Samsung is expected to handle about 10 per cent of Nvidia's AI GPU packaging, according to people familiar with the matter. But they added that Samsung must meet Nvidia's requirements and pass its quality tests.

He also mentioned that TSMC is planning to expand its 2.5D packaging capacity by more than 40 per cent to meet Nvidia's needs, meaning Samsung may not be able to strike a deal unless it quickly passes Nvidia's assessment.

It is worth mentioning that SK Hynix is the only company in the world that can produce HBM3 chips on a large scale, so other manufacturers can only find it if they want to buy HBM3E.

However, Samsung plans to start producing HBM3 chips later this year. Kyung Kye-hyun, head of Samsung Semiconductor, previously said that the company's HBM3 was rated as excellent by customers, and he expected HBM3 and HBM3P to contribute to profit growth in the chip division from next year.

In addition, Samsung is committed to developing a bump-free package for high-level HBM by 2025, which helps reduce package height, also known as copper-copper direct bonding or hybrid bonding (Cu-Cu Hybrid Bonding).

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.