According to a DigiTimes story, demand for advanced compute GPUs such as Nvidia’s A100 and H100 is so great that TSMC is now considering to expand its advanced packaging capacities. TSMC has reportedly promised to produce an additional 10,000 CoWoS wafers for Nvidia through 2023. Given that Nvidia receives approximately 60-ish A100/H100 GPUs each wafer, it would imply an additional 600,000 top-end data centre GPUs.
The predictions indicate a monthly rise of 1,000 to 2,000 wafers for the balance of the year. TSMC’s monthly CoWoS output ranges between 8,000 and 9,000 wafers, therefore supplying Nvidia with an additional 1,000 to 2,000 wafers monthly will greatly increase TSMC’s high-end packaging facility utilisation rate.
Due to the increased demand for CoWoS services, this rise may result in a supply shortage for other industry participants, which is why TSMC is reportedly planning to expand its advanced packaging capacities.
TSMC’s output boost is purportedly intended to meet the increasing demand for Nvidia’s AI processors, which are widely used in the sector.
Its similar to how Google has just unveiled their new A3 supercomputer, which is built on Nvidia’s H100 and has 26 ExaFLOPS of AI performance. Similarly, other big companies, including Microsoft, Oracle, and Elon Musk’s future AI startup, have purchased tens of thousands of Nvidia’s AI GPUs in recent months.
It is unclear which compute GPUs Nvidia aims to increase production of, given its current lineup comprises the A100, A30, H100, and China-exclusive A800 and H800 GPUs.
Also Read:
