Google (GOOGL, Financial) has introduced its seventh-generation TPU, Ironwood, designed for inference with performance 3,600 times higher than its first generation. It competes head-to-head with NVIDIA's (NVDA) B200, marking a significant advancement in AI hardware. Ironwood stands as Google's most powerful and scalable custom AI accelerator, specifically targeting inference tasks, and will be launched later this year. Its efficiency is highlighted by a 29-fold improvement compared to the initial TPU and offers performance 24 times that of the world's largest supercomputer.
Compared to the sixth-generation TPU Trillium, Ironwood doubles the power efficiency and outperforms the 2018 Cloud TPU by nearly 30 times. Google's advanced liquid cooling and optimized chip design ensure reliable performance under continuous AI workloads. Performance comparisons by OpenAI researchers indicate Ironwood's capabilities rival and even slightly surpass NVIDIA's GB200, combining impressive performance with power efficiency for cost-effective AI workload operations.
Additional updates from Google's cloud conference include advancements in AI platforms like Veo 2 and the introduction of Lyria, a text-to-music model, under the new Vertex AI, which supports video, image, speech, and music. The conference also highlighted Gemini Code Assist, an AI coding assistant, now capable of intelligent functions within programming tasks such as building applications from product specs in Google Docs and translating code between languages.
