Google today at the Google Cloud Next event, announced the upcoming eighth-generation TPU chips with two main purposes, namely for training purposes and also for making inference. Google named these chips as TPU 8t and TPU 8i respectively.
Google says that TPU has been powering Gemini for several years, and with this new chip, it is more powerful and efficient, including for training, offering services, and for AI agent tasks.
The TPU 8t, which focuses on training AI models, is said to bring almost 3 times the computing power per pod compared to the previous generation, thus allowing model development time to be accelerated from months to just a few weeks.
Meanwhile, for the purpose of processing the use of AI models, the inference-focused chip, TPU 8i plays a key role, including for AI agents. This chip is also said to offer better performance, and is also optimized for complex tasks and larger memory support.
Google says that it is co-designed for Gemini, in addition to being open to various leading platforms on the market. Previously, it was said that several parties, including Meta, were looking at using this Google-produced chip, which is seen as more efficient and economical than other providers.
