"GPU turning point" coming in the first quarter of next year? Faster than market expectations by one to two quarters!
The GPU production capacity is about to reach a turning point, and the cooperation between TSMC and Samsung is expected to alleviate NVIDIA's supply difficulties. The supply chain problem of HBM3 will be alleviated, which is currently the most advanced memory chip technology, mainly used for NVIDIA's flagship GPU H100. In addition, TSMC, NVIDIA's advanced packaging supplier, will also expand its production capacity. Design changes and order cancellations by cloud service providers are also important factors in the turning point of GPU production capacity. It is expected that after the advanced packaging capacity problem is resolved, the supply of related components will gradually increase.
Under the joint efforts of TSMC, Samsung, and other manufacturers, Nvidia's bottleneck problem is expected to be quickly resolved, and GPU production capacity is about to reach a turning point.
According to Taiwan's Economic Daily News, Samsung will join Nvidia's HBM3 supply chain to alleviate the supply problem caused by SK Hynix's monopoly in the past. HBM3 is currently the most advanced memory chip technology, mainly used for Nvidia's flagship GPU, H100.
In addition, TSMC, Nvidia's advanced packaging supplier, is further expanding its advanced packaging capacity through equipment changes and strengthening external cooperation. Combined with factors such as changes in external demand, design adjustments, and order changes, TSMC is expected to break through the bottleneck of advanced packaging capacity in the first quarter of next year, one quarter to half a year earlier than industry expectations.
In September, Wall Street News mentioned that TSMC believed at the time that the shortage of CoWoS packaging supply would take about 18 months to alleviate, which affected the supply of Nvidia GPUs. Now, the shortage of advanced packaging capacity is expected to be resolved ahead of schedule, which means that the supply problem of Nvidia AI chips will also be alleviated.
In a previous earnings conference call, Nvidia confirmed that it has certified other CoWoS advanced packaging suppliers as backups. According to the Economic Times, certifying the front-end and back-end capacity of other CoWoS advanced packaging suppliers will help TSMC and its partners achieve a monthly CoWoS capacity of about 40,000 wafers in the first quarter of next year.
In a recent press release, Samsung stated that it has been closely cooperating with TSMC in the previous generations of HBM technology, supporting the compatibility of CoWoS process and the interconnectivity of HBM. After Samsung's memory chip division joins TSMC's OIP 3DFabric Alliance in 2022, it will further expand its scope of work and provide solutions for future generations of HBM.
In addition to the expansion of production by the two major suppliers, a key factor in the upcoming turning point in GPU production capacity is the gradual design changes by Nvidia's major customers, namely cloud service providers, which reduce the usage of advanced packaging, as well as the deletion of orders by some duplicate customers.
Industry insiders predict that after the alleviation of the advanced packaging capacity problem, it is expected to take about three months for the impact to be transmitted from the semiconductor upstream to the downstream system and component, and it is expected to benefit the gradual increase in shipments of related supporting component manufacturers.