NVIDIA’s AI Graphics Cards in High Demand, Production Capacity to Increase by 30%
Due to the surging popularity of artificial intelligence (AI), NVIDIA’s AI graphics cards are currently facing a shortage in supply. Among them, the high-end A100 and H100 models are especially sought after, leading manufacturers to place large orders. In response, NVIDIA is expediting production and has sought foreign aid to increase its manufacturing capacity by 30%.
Unlike the capacity bottleneck experienced by other GPU manufacturers, NVIDIA’s high-end GPU utilizes TSMC’s 7nm and 5nm process for foundry, but the limitation lies in the CoWos (Chip-on-wafer-on-substrate) advanced packaging used. Previously, NVIDIA relied on TSMC for packaging, but the company’s capacity increase is restricted and it also has orders from other firms such as AMD, resulting in a struggle to secure sufficient production capacity.
To tackle this issue, UMC has stepped in to assist NVIDIA. Beginning this July, UMC will increase the production capacity of CoWos interposers by an additional 3,000 wafers per month. Furthermore, packaging and testing company Amkor Packaging will also collaborate with the two companies. This joint effort is expected to boost the overall production capacity of A100 and H100 by 30%.
With the increase in supply, it is anticipated that the shortage of A100, H100, and other graphics cards will be alleviated. The situation of price increases and panic buying is expected to be resolved in the second half of the year.
The H100 model utilizes the GH100 GPU core, a customized version of TSMC’s 4nm process manufacturing. It features 800 transistors, integrates 18,432 CUDA cores, 576 Tensor cores, 60MB L2 cache, supports 6144-bit HBM (high-bandwidth memory), and is compatible with PCIe 5.0.
In terms of performance, the H100 boasts FP64/FP32 speeds of 60TFlops (60 trillion calculations per second), FP16 speeds of 2000TFlops, and TF32 speeds of 1000TFlops – three times that of the A100. Additionally, it delivers FP8 speeds of 4000TFlops, offering six times the performance of the A100 model.
Sources: NVIDIA, UMC, TSMC
Further reading: