Samsung is in talks to manufacture a key component for Google's tenth generation of AI accelerators, codenamed Icefish, marking Google's shift toward a supply diversification strategy.
What Happened
Samsung plans to use its 2nm process to manufacture the I/O die, which is responsible for connecting the processor to high-speed HBM memory. Meanwhile, the main computing cores for the Icefish project are planned to be ordered from TSMC using 1.4nm technology. Mass production of this architecture is scheduled for 2028.
Context
The Icefish project demonstrates a move away from monolithic designs toward heterogeneous manufacturing using chiplet architecture. This allows components manufactured at different foundries and using different process nodes to be combined within a single high-performance device.
Why It Matters for the Industry
Google is implementing a multi-foundry approach to reduce dependence on TSMC's capacity and diversify supply risks. For Samsung, this is an opportunity to regain market share in advanced process nodes and strengthen competition for contracts to produce auxiliary components, such as I/O dies and interposers.
Why It Matters for Users
The transition to complex multi-die systems is changing the rules of the game in the race for AI chip performance. This confirms a long-term industrial trend toward increasing the complexity of the physical computing layer and moving from monolithic chips to modular, optimized systems.
Sources
Author
Look at AI, Editorial Team