July 3, 2023 – The global AI GPU market is witnessing a dominant presence by NVIDIA, with a market share of over 90%, turning its chip manufacturing rights into a fiercely contested battleground for vendors.
NVIDIA, renowned for its GPUs used in ChatGPT, currently relies exclusively on TSMC for the production of its flagship chips, the A100 and H100 GPUs. In response to overwhelming demand for these chips, TSMC had already made the decision in early June to expand its packaging capacity at NVIDIA’s request.
TSMC’s ability to exclusively manufacture NVIDIA chips can be attributed to its advanced packaging technology known as CoWoS. With recent advancements in semiconductor manufacturing processes reaching a thickness as thin as a human hair, the significance of packaging technology as a means to enhance semiconductor performance has become increasingly prominent.
During the packaging process, chips are stacked in a 3D configuration, reducing the distance between them and thereby enabling faster interconnectivity. This packaging approach can yield substantial performance improvements of up to 50% or even more.
TSMC first introduced CoWoS technology in 2012 and has continuously upgraded its packaging capabilities since then. Today, flagship products from NVIDIA, Apple, and AMD heavily rely on TSMC’s advanced packaging technology. This also explains why Samsung Electronics, despite taking the lead over TSMC in achieving mass production of 3nm chips in 2022, still sees NVIDIA, Apple, and other industry giants opting for TSMC’s production lines.
In an effort to surpass TSMC’s CoWoS, Samsung is currently developing more advanced packaging technologies such as I-cube and X-cube. Furthermore, there are reports indicating Samsung’s focus on 3D packaging, wherein multiple chips are vertically stacked to improve performance. An industry insider stated, “Samsung and TSMC are on track for a head-on clash in the field of packaging very soon.”