Google unveils two new TPUs designed for the “agentic era”

In a bold move to redefine AI infrastructure, Google has introduced its latest Tensor Processing Units (TPUs), marking a significant leap forward from its previous generation. While many tech companies are scrambling to acquire Nvidia's AI accelerators, Google is leaning heavily on its custom-built TPUs, which are tailored specifically for their cloud AI services. The newly revealed eighth-generation TPUs, named TPU 8t and TPU 8i, are designed to cater to distinct phases of AI development. The TPU 8t focuses on training, aimed at drastically reducing the time it takes to train advanced AI models from several months to just weeks. This is a crucial advancement in the AI lifecycle, as effective training is essential before deploying models for tasks like data analysis or creative content generation. Google’s updated server clusters, referred to as “pods,” now boast an impressive configuration of 9,600 chips alongside two petabytes of shared high-bandwidth memory. This setup enables the TPU 8t to scale linearly, potentially accommodating up to a million chips within a single logical cluster. Such innovations not only enhance the speed of training large AI models but also contribute to rising RAM prices, impacting the broader tech landscape. With a remarkable computational capability of 121 FP4 EFlops per pod, the TPU 8t is nearly three times more powerful than its predecessor, the Ironwood TPU. As Google positions itself at the forefront of the AI revolution, these advancements underscore the company’s commitment to shaping the future of artificial intelligence.