TSMC admitted in mid-2023 that demand for its chip-on-wafer (cowos) has outstripped its capacity, and the company has vowed to double its capacity by the end of 2024. According to Moneyudn.com reported that Nvidia wants to ship as many of its high-demand AI processors as possible, which is why it is leveraging Intel's advanced packaging technology (in addition to TSMC's). As with all unconfirmed reports, we must take this with a grain of salt until the companies in question comment. The deal allegedly produces 5,000 wafers per month, which is equivalent to 300,000 of NVIDIA's H100 chips per month based on a quick calculation on the back of the napkin (assuming the yield is perfect and the contract is for the H100).
TSMC is expected to remain the main player, providing about 90% of NVIDIA's advanced packaging capacity. But starting in the second quarter, Nvidia is also preparing to use Intel's production capacity for at least some of its products, the report said. If this information is accurate, increasing Intel's capacity will allow NVIDIA to produce more GPUs for AI and HPC workloads, allowing it to meet demand for its existing products more quickly. However, there is a catch.
All of NVIDIA's current and previous generation products, such as the A100, A800, A30, H100, H800, H200, and GH200, feature TSMC's CODOS-S packaging process, which relies on a silicon interposer. The closest advanced packaging technology that Intel has is called Foveros, which also relies on an interposer, albeit a different kind of interposer (Codos-S may use a 65nm interposer, while Foveros uses a 22FFL interposer).
To use Intel's FoverOS, NVIDIA needs to validate the technology and then qualify actual products, which may have slightly different characteristics than TSMC's packaged products (because the interposers are manufactured using different process technologies and have different bump pitches), so the company's partners may also need to verify their qualifications before deploying. If that's the case, it will be interesting to see if Nvidia outsources only part of its products to Intel, or all of them.
Translation**: