AI performance benchmark reports and publications

Google Unveils Gemini 2.0 and Trillium TPUs: Faster AI Training and Multimodal Capabilities

Dec 12, 2024, 10:31 PM

Google has announced the launch of Gemini 2.0, its new AI model featuring advanced multimodal capabilities, including the ability to generate images and audio in multiple languages, and edit images using text commands. The Gemini 2.0 family includes a version called Gemini 2.0 Flash. Users have noted that Gemini 2.0 demonstrates impressive speed and can process long texts, providing detailed analyses; for instance, it can read an entire book and accurately answer questions about its content. Alongside the AI model, Google has made its sixth-generation Trillium TPUs generally available; these AI chips are designed to power Gemini 2.0 and offer improved performance, with 4x faster AI training and 3x better inference throughput compared to the previous generation. The Trillium TPUs achieve significant scaling efficiency, reaching 99% with 3,072 chips and 94% with 6,144 chips for training. Google has also introduced a unified interface for Gemini Developer APIs in Google AI Studio and Vertex AI APIs, enhancing accessibility for developers and allowing evaluation and experimentation with the model.