Nvidia reportedly mulls outsourcing partial AI GPU production to Samsung

Jessica Tsai, Taipei; Willis Ke, DIGITIMES Asia 0

Nvidia CEO Jensen Huang. Credit: AFP

Nvidia is reportedly mulling outsourcing a portion of its AI GPUs to Samsung Electronics for fabrication amid increasingly tight capacity supply from TSMC. Industry observers noted that Samsung could potentially secure some orders from Nvidia if its 3nm trial products pass performance validation and its 2.5D advanced packaging technology meets the requirements of the US chipmaker.

South Korean media Chosun Biz, citing local semiconductor industry sources, recently reported that Nvidia is in talks with Samsung regarding contract production of chips, with their performance validation discussions based on the most advanced processes.

Most industry insiders in South Korea believe that the possibility of Samsung securing large-scale orders from Nvidia is not high, but the US chipmaker may not rule out enlisting Samsung as a secondary foundry partner due to the challenges of relying solely on TSMC to fulfill all its AI GPU orders from customers.

With the rise of AI-related services, IT companies are actively procuring GPUs for establishing AI data centers. A Market Watch estimate shows that the global GPU market scale may grow to around US$33.463 billion by 2028 from US$19.711 billion in 2021, representing a CAGR of 7.85% during the period.

At the moment, the AI GPU market is mainly dominated by Nvidia and AMD, but in terms of AI GPUs for driving large language models (LLM) like ChatGPT, Nvidia boasts a market share of over 90%.

Some industry analysts said that in response to the rapid surge in demand for AI GPUs, Nvidia may collaborate with Samsung to mitigate the risks of supply shortages, although the South Korean maker's 3nm manufacturing process has yet to achieve stable yield performance. Therefore, in addition to obtaining performance validation for its 3nm test chips, whether Samsung's advanced packaging technology can meet Nvidia's requirements is another key factor in determining potential collaboration between both parties.

In an effort to upgrade its packaging technology, Samsung in June just established the Multi Die Integration Alliance (MDI) seeking to expand the ecosystem of collaboration partners, continuously improve stacking technology and actively invest in 2.5D and 3D packaging technologies. In contrast, TSMC has developed its 2.5D CoWoS packaging technology for many years, successfully utilizing the technology to improve computing performance of chips, and has thus become a top-choice foundry partner for AI and HPC chipmakers.

In the past, Samsung had also fulfilled GPU orders from Nvidia, including producing GeForce RTX30 series using 8nm process in 2020 and fabricating 14nm mobile GPU chips before that. However, Nvidia has outsourced the production to TSMC for all its current flagship A100 and H100 GPUs supporting ChatGPT and other AI applications.