Nvidia plans to increase data center GPU production in 2024
Recently, NVIDIA unveiled its financial results for the second quarter of the 2024 fiscal year. Bolstered by surging demands for High-Performance Computing (HPC) and Artificial Intelligence (AI) over the past few months, quarterly revenue surpassed $10 billion for the first time, reaching an impressive sum of $13.507 billion. Among these figures, the revenue from the data center business amounted to $10.32 billion, eclipsing the $2.49 billion from the gaming sector. These two once evenly-matched pillars of NVIDIA’s revenue stream now demonstrate a widening disparity in their respective earnings.
As per the Financial Times, NVIDIA has set in motion plans to bolster the production of its data center GPUs for 2024, aiming to meet the robust market demand for products like the A100, H100, and other computational CPU/GPUs. Such an endeavor hints at potentially astronomical revenue figures in the offing. Insiders reveal that NVIDIA is considering amplifying the production of its GH100 units from the current 500,000 to an ambitious range of 1.5 to 2 million.
Products leveraging the GH100 chip include the H100 computational card and the GH200 Grace Hopper, among others. Escalating the supply of these premium products is no facile undertaking. Every facet of the supply chain is implicated, as evidenced by NVIDIA’s recent concerns with TSMC over the CoWoS packaging capacity. The intrinsic complexity in the design of the GH100 chip makes its mass production a challenging endeavor. To drastically augment production levels, several bottlenecks must be overcome.
Primarily, to ensure the desired production of GH100, NVIDIA requires TSMC to enhance its 4N process capacity. Preliminary estimates suggest that a single 300mm wafer can produce up to 65 GH100 chips. Should NVIDIA’s ambition to produce 2 million units materialize, approximately 31,000 wafers would be necessary—a feasible task given TSMC’s monthly capacity of 150,000 units at the 5nm process node.
Furthermore, NVIDIA’s reliance on TSMC’s CoWoS packaging means the current capacity is severely lacking. This deficiency partially explains why NVIDIA contemplates diversifying its packaging orders, potentially entrusting a portion to Samsung.
Additionally, the GH100 demands HBM2E, HBM3, and HBM3E. To meet this need, NVIDIA must secure a substantial supply of HBM, possibly sourcing from giants such as Samsung, SK Hynix, and Micron concurrently.
Lastly, the onus lies on NVIDIA’s partners to integrate products predicated on the GH100 chip into servers. This not only gauges the partners’ throughput but also underscores the perennial necessity to maintain substantial market demand.
Should all these conditions coalesce seamlessly, NVIDIA is poised to reap even more staggering profits in the ensuing year.
