Nvidia Explains GPU Supply Issues
Over the past several months, a fervor for artificial intelligence (AI) tools, spearheaded by ChatGPT, has swept across the globe, leading to a significant surge in demand for data center GPUs such as NVIDIA’s A100 and H100. Currently, these data center GPUs, including NVIDIA’s A100 and H100, are manufactured and packaged by Taiwan Semiconductor Manufacturing Company (TSMC), with SK Hynix supplying the HBM3 chips. In order to meet the market’s demand for data center GPUs, TSMC has urgently ordered new packaging equipment, aiming to expand 2.5D CoWoS packaging capacity by over 40%.
Recently, Elon Musk noted that the market demand for data center GPUs is disproportionately greater than the available supply, acknowledging a substantial discrepancy, even remarking that these GPUs are “even harder to purchase than white powder.” According to Computerbase, Charlie Boyle, the Vice President and General Manager of DGX systems, clarified that the issue does not stem from NVIDIA’s miscalculation of market demand, nor from problems with TSMC’s chip manufacturing output or quality yield, but rather from the capacity of 2.5D CoWoS packaging.
2.5D CoWoS packaging is a multi-step, high-precision engineering process, the complexity of which reduces the quantity of GPUs that can be packaged within a specific timeframe, thereby directly impacting the ultimate supply chain. Currently, the supply bottleneck caused by the 2.5D CoWoS packaging capacity for data center GPUs appears to be more severe than anticipated. TSMC has indicated that it may take approximately a year and a half for the packaging process to return to normalcy. This signifies that NVIDIA must make compromises, as there is insufficient time and capability to package all the products.
NVIDIA has taken the lead in the AI market, and TSMC is among the few companies possessing high-performance packaging technology. Coupled with their leading-edge semiconductor manufacturing processes, this combination is virtually insurmountable in the current AI domain. Samsung, ranking second in the wafer foundry market, is not only trailing TSMC in semiconductor manufacturing technology but also in packaging techniques. Presently, Samsung is increasing its investments in a concerted effort to narrow the gap between the two entities.