Cloud Next ‘26: All About the New Eighth-Generation Google TPUs

At the Cloud Next ‘26 technology conference, Google unveiled the eighth generation of its custom Tensor Processing Units (TPUs). Developed in close partnership with Google DeepMind, these new chips are engineered for maximum energy efficiency.

About TPU 8t and TPU 8i

In this generation, Google has focused on narrow component specialization, introducing two processor versions: TPU 8t (Training) and TPU 8i (Inference). 

• TPU 8t (Training): This chip is designed for training large-scale neural networks. Its architecture allows for the scaling of up to 9,600 chips into a single supermodule, providing access to a massive 2 PB of memory with HBM bandwidth. 
• TPU 8i (Inference): This chip is optimized for executing pre-trained algorithms (inference) and delivers 80% higher performance compared to its predecessors. Consequently, the chip ensures zero latency and rapid response times for AI services.

Advancements in Energy Efficiency

While the new generation of chips surpasses previous solutions like Ironwood in power, it also demands more from data center power grids. To enhance efficiency, Google integrated an adaptive power management system into the TPU 8t and 8i. This allows the chips to adjust resource consumption in real-time based on task complexity.

Partnership with NVIDIA

Google continues to collaborate with GPU market leaders. The company confirmed it will be among the first to offer its customers access to the latest NVIDIA Vera Rubin NVL72 systems.