From Tom's Hardware: TSMC is negotiating with Nvidia to produce its Blackwell GPUs for AI at its Fab 21 near Phoenix, Arizona, reports Reuters. If the companies ink an appropriate agreement, Nvidia might get additional Blackwell GPUs produced in the U.S., undoubtedly pleasing its customers and politicians. There are a couple of catches, though.
Nvidia's Blackwell GPUs are made using TSMC's custom 4NP (4nm-class) process technology that can be relatively easily ported from the foundry's fabs in Taiwan to its Fab 21 in Arizona, which is designed to make chips on 4nm and 5nm-class production nodes.
Nvidia's highest-end Blackwell GPUs for AI and HPC rely on TSMC's CoWoS-L (B100, B200, B300, etc.) and CoWoS-S (B200A, B300A) packaging. TSMC does not have an advanced packaging facility in the U.S., and its partner Amkor will only start packaging chips in Arizona in 2027. As a result, Blackwell AI silicon produced in Arizona will need to be shipped back to Taiwan for final assembly, as all of TSMC's CoWoS packaging capacity remains in Taiwan. Shipping to Taiwan would add costs, but since most AI servers are assembled in Taiwan anyway, this should not create a significant logistics problem.
View: Full Article
