Google Introduces Two New TPUs for the Age of Autonomous Systems
agents chips google inference tpu training
| Source: Mastodon | Original article
Google debuts two new TPUs to accelerate AI model training. They reduce training time from months to weeks.
Google has unveiled two new Tensor Processing Units, the TPU 8t and TPU 8i, designed for the "agentic era". As we reported on April 22, the company has been working on its eighth-generation TPUs, and these new chips reflect Google's belief that AI agents require fundamentally different architectures. The TPU 8t handles training, while the TPU 8i handles inference, splitting the workload to reduce frontier model training time from months to weeks.
This development matters because it signals Google's commitment to advancing AI infrastructure, particularly in the area of agentic AI, which involves autonomous agents that can learn and adapt. By reducing training times, Google aims to make it easier for developers to build and deploy AI models, potentially leading to breakthroughs in areas like robotics and natural language processing.
What to watch next is how these new TPUs will be integrated into Google's cloud services, including the updated Tensor 8t server clusters, which will house 9600 chips with two petabytes of shared high-bandwidth memory. With the TPU 8t and TPU 8i set to be available later this year, developers and researchers will be keen to see how these new chips perform in real-world applications, and how they compare to competing offerings from companies like Nvidia.
Sources
Back to AIPULSEN