Nvidia has reportedly shifted some orders to Taiwan Semiconductor Manufacturing Co. from GeForce RTX 4090 over to H100 compute processors based on the Hopper architecture. While the information comes from Mydrivers and we could not verify it at press time, there are factors that support the report. It could be a result of the U.S. sanctions of China's supercomputer sector.
Nvidia’s latest gaming and high performance computing (HPC) chips are made using TSMC's customized process technology called 4N, as opposed to generally used N4 (a 4nm-class node that belongs to TSMC's N5 family of fabrication processes). It should be noted that orders to TSMC will not bring an immediate effect on anything given how long it takes to make a modern processor and how ordered are placed. Meanwhile, the information has not been confirmed either directly or indirectly, so take it with a grain of salt.
At the moment Nvidia uses TSMC to produce a host of its GPUs of different complexity, which includes its lates GH100 (80 billion transistors, 814mm^2), AD102 (76.3 billion transistors, 608mm^2), AD103, and AD104 processors. Refocusing allocation at TSMC from one product may not sound too difficult, but there are nuances. Nvidia's AD102 GPU powers the GeForce RTX 4090, which tops our list of the best graphics cards for gaming as of October 2022.
From TSMC's perspective, it does not matter what to produce on its N5-capable production lines. Since both GH (Hopper) and GA (Ada Lovelace) are made on the same fab(s), the foundry has (almost) no problems shifting production within given allocation. Meanwhile, Nvidia also uses TSMC's advanced packaging capacities to make its Hopper GPUs for compute (i.e., it uses one TSMC's 3D Fabric 2.5D technologies to connect HBM3 memory to the chip itself).
From Nvidia's perspective, the situation is quite a bit different. Shifting TSMC's allocation, an uneasy thing to get to on the first place, is one thing. The chipmaker is prioritizing a product that you sell for thousands (Nvidia's H100 products cost $10,000 a piece, whereas consumer-oriented GeForce RTX 4090 carries a recommended price tag of $1,499). While complexity of the AD102 and GH100 is comparable, earnings and margins that Nvidia gets for these products are quite different, which eventually has an effect on the company's research and development buget.
While we all know that demand for graphics cards from gamers is high, the production shift might be a direct result of the U.S. sanctions against China's supercomputer sector, which might have cost Nvidia some $400 million should it not get a waiver to ship compute GPUs to Chinese entities.
