AMD is set to produce high-performance chips at TSMC’s new Arizona fab, making it the second high-profile client for the fab after Apple. Independent journalist Tim Culpan reports that sources close to the matter confirmed the agreement today, though TSMC declined to comment.
TSMC Fab 21, located near Phoenix, Arizona, has begun trial production on its 5nm node, a process node family that includes the N4/N4P/N4X and N5/N5P/N5X processes. While its phase 1 production has not fully commenced, the Apple A16 Bionic chip is currently in production at Fab 21 using the N4P process. The A16 Bionic has been around since mid-2022 and is an excellent manufacturing test for the young fab, which reports “small, but significant” numbers of the chips currently being produced for various Apple products. Fab 21’s current yields were described as similar to TSMC’s Taiwan fabs by Bloomberg last month.
However, the chip(s) that AMD plans to produce at Fab 21 is currently unknown. According to sources, production is currently in planning, with tape-out and manufacturing of the chip both set to happen in Arizona starting next year. Fab 21’s Phase 1 is limited to the N4 and N5 technologies, barring the possibility of any consumer chips newer than RDNA 3 and Zen 4.
AMD’s enterprise AI chips in the CDNA 3 family powering Instinct MI300-series accelerators are a likely suspect. The MI325X, due to release in Q4 2024, is on the N4 node, while the upcoming MI350 will be on TSMC’s N3. Arizona could be the place where the MI325X is manufactured after its initial production waves. It must be said that this is a speculative take; AMD could decide to produce a yet-unannounced AI or mobile chip at Fab 21 instead.
AMD’s HPC chips that are manufactured in Arizona must, at first, be shipped overseas for packaging. However, the recent agreement between Amkor and TSMC to team up for advanced packaging in Arizona will more firmly entrench an AI chip supply chain in the U.S. Amkor’s $2 billion Arizona chip test and packaging facility, currently under construction and expected to begin production as soon as 2026, will be permitted to use TSMC’s patented CoWoS and InFO packaging technologies, allowing AI and HPC chips to be packaged more completely in the U.S. GPUs specifically rely on CoWoS tech to connect with their high-bandwidth memory (HBM) chips. Amkor began construction on its Arizona fab counting on TSMC to become its major customer, but the agreement to work with the patented CoWoS technology is still a happy surprise for many.
2 Likes