TSMC will increase advanced packaging capacity

In recent years, Artificial Intelligence (AI), High-Performance Computing (HPC), and personal computing have been the impetus behind the advancement of cutting-edge processes and packaging technologies. The demand in the market is burgeoning rapidly. As a frontrunner in the semiconductor foundry industry, Taiwan Semiconductor Manufacturing Company (TSMC) persistently amplifies its capacity for these advanced processes and packaging to accommodate market fluctuations. In recent months, a global frenzy has been incited by AI tools, primarily ChatGPT, which has substantially elevated the demand for data center GPUs such as Nvidia’s A100 and H100.

TSMC CFET Transistor

According to DigiTimes, TSMC is presently strategizing to expand its advanced packaging capacity, urgently procuring new packaging equipment to fulfill order requirements within this year. As TSMC and Nvidia both previously underestimated the market’s need for data center GPUs, the existing packaging equipment is no longer adequate.

It is understood that the current demand for packaging technologies like CoWoS far surpasses the existing capacity. TSMC has assured Nvidia that an additional 10,000 CoWoS wafers will be processed during 2023. Given that approximately 60 A100/H100 GPU chips are manufactured per wafer, this suggests an increase of around 600,000 data center GPU orders. TSMC plans to augment its CoWoS wafer count by 1,000 to 2,000 each month for the remainder of this year, with a monthly output estimated to be between 8,000 to 9,000 wafers, thereby increasing the utilization of its advanced packaging facilities.

There have been prior reports stating that the delivery cycle for high-performance computing GPUs has been extended from three months to six months, and in certain cases, a longer wait might be required. Some new orders are projected to be fulfilled only by December this year, indicating a waiting period exceeding six months. Currently, TSMC’s 4/5nm and 6/7nm capacity utilization rates have essentially reached saturation to cope with Nvidia’s new orders, which include the A100, A30, H100, A800, and H800 computation cards.