Google Icefish TPU to Use Samsung 2nm and TSMC 1.4nm Process Nodes

Google plans to manufacture its next-generation Tensor Processing Unit using a split foundry strategy that combines Samsung and TSMC process technologies. The chip, internally known as Icefish, represents the tenth generation of Google's custom AI silicon designed for cloud data centers. This architectural shift aims to address global semiconductor capacity constraints while pushing performance boundaries for upcoming hardware platforms.

Hybrid manufacturing approach leverages specialized foundry strengths

The new Tensor Processing Unit targets improved inference performance and operational energy efficiency for Nvidia's Vera Rubin platform. Google collaborates with MediaTek on the design phase, though the manufacturing execution relies on external foundry partners. The project remains in active development, which means the final specifications could change before mass production begins.

Samsung will fabricate the input/output dies using its 2-nanometer process node. TSMC handles the core compute engine portion of the chip with a more advanced 1.4-nanometer manufacturing technique. This hybrid approach allows Google to leverage specialized strengths from both semiconductor manufacturers for different sections of the processor.

The chip design focuses on optimizing energy consumption during AI workloads rather than raw peak performance metrics alone. Google intends to deploy this hardware across its global cloud infrastructure to support machine learning tasks. The split manufacturing process requires careful integration between components produced at separate facilities.

Mass production for Icefish is expected as early as 2028, according to current project timelines. The device will launch globally without specific regional restrictions mentioned in available reports. Google continues to refine the design while managing supply chain dependencies across multiple foundries.

Google has previously developed nine generations of Tensor Processing Units before this latest iteration. The company maintains a long-term strategy for custom silicon development to reduce reliance on third-party GPUs for cloud computing tasks. This new chip represents a continuation of that internal hardware roadmap.