According to reports, Nvidia has moved certain orders to Taiwan Semiconductor Manufacturing Co. from GeForce RTX 4090 to Hopper-based H100 computing units. While the information originates from Mydrivers and we were unable to verify it at the time of publication, there are several aspects that substantiate the story. It might be the outcome of US sanctions against China’s supercomputer sector.
Nvidia’s newest gaming and high performance computing (HPC) processors are manufactured utilizing TSMC’s proprietary process technology known as 4N, rather than the more commonly used N4 (a 4nm-class node that belongs to TSMC’s N5 family of manufacturing processes).
It should be emphasized that orders to TSMC will not have an immediate impact due to the time it takes to manufacture a contemporary CPU and how orders are placed. Meanwhile, no one has corroborated the information, either directly or indirectly, so take it with a grain of salt.
It makes no difference to TSMC what it produces on its N5-capable production lines. Because both GH (Hopper) and GA (Ada Lovelace) are manufactured on the same fab(s), the foundry has (almost) no difficulty moving production within the allotted allocation. Meanwhile, Nvidia makes its Hopper GPUs for computation using TSMC’s sophisticated packaging capabilities (i.e., it employs one of TSMC’s 3D Fabric 2.5D technologies to connect HBM3 memory to the chip itself).
The scenario is significantly different from Nvidia’s point of view. Shifting TSMC’s allocation, which is a difficult thing to do in the first place, is one thing. The chipmaker is favoring a product that you sell for thousands of dollars (Nvidia’s H100 devices cost $10,000 each, whilst the GeForce RTX 4090 is suggested for $1,499). While the complexity of the AD102 and GH100 is equivalent, Nvidia’s earnings and margins for these devices are substantially different, which has an impact on the company’s R&D budget.
While we all know that gamers have a strong demand for graphics cards, the manufacturing move might be a direct result of US sanctions against China’s supercomputer industry, which could have cost Nvidia $400 million if it did not receive a waiver to sell compute GPUs to Chinese businesses.
