TSMC: Shortage of AI accelerators will continue until the end of 2024

According to TSMC Chairman Mark Liu, it will take some time before the manufacturer can meet the high demand for AI accelerators. The Tawainese chip maker, which is the sole supplier used to manufacture Nvidia’s H100 and A100 chips, expects that the pressure on supplies will only decrease around the end of 2024.

Liu indicates that the shortage is not caused by a lack of chips, but by a shortage of packaging capacity. For example, in the advanced accelerators of team green, a compute chip is combined with various HBM dies (memory). To place this computing power on the same substrate, a special technique is required, which TSMC calls Chip on Wafer on Substrate (CoWoS).

The ‘GPU’ part of the Grace Hopper Superchip (right) is paired with six HBM3e stacks, giving an available capacity of 141 GB.

As demand for CoWoS has tripled in one year, the company is currently only able to meet 80% of demand. We are working hard to set up more production resources, but this will take time. TSMC recently announced that it will double its CoWoS capacity by the end of 2024, which matches the Chairman’s aforementioned estimated time window.