Particle.news
Download on the App Store

Google Reportedly in Talks With Samsung to Build Part of Its Next TPU

The talks would let Google split production across foundries to lock in capacity for growing demand for AI chips.

Overview

  • Multiple outlets reported on June 12 that Google is in preliminary talks with Samsung Foundry to make a memory input/output die for its tenth‑generation TPU, codenamed Icefish, using Samsung’s 2nm process.
  • The same reports say Taiwan Semiconductor Manufacturing Co. would still produce Icefish’s main compute die while the Samsung part would link the compute die to high‑bandwidth memory for large AI workloads.
  • Google is reported to be working with MediaTek on Icefish’s design and is separately discussing large‑volume production with Intel for possible mass output in 2028, though no contracts have been announced.
  • Splitting the chip into a compute die and a memory I/O die is a technical choice to balance yield and performance when using different process nodes and packaging for HBM connections.
  • If talks proceed, the move would mark a shift from Google’s long reliance on TSMC and reflect industry pressure on foundry capacity given Samsung’s recent 2nm wins and broader competition among TSMC, Samsung, and Intel.