Google unveils TPU 7th Gen Ironwood with class-leading capacity of 42.5 ExaFLOPS, citing TPU performance that makes selling Gemini more cost-effective than OpenAI and DeepSeek.

At the Google Cloud Next 2025 event, Google unveiled the TPU Ironwood chip, their in-house AI chip succeeding the TPU v5p. This chip highlights the Inter-Chip Interconnect (ICI) for creating large-scale clusters of nearly ten thousand chips. The Ironwood TPU boasts a processing power of 4,614 TeraFLOPS, exceeding the v5p by more than ten times (outperforming the Trillium by 4.7 times). Each chip comes equipped with 192GB of HBM memory and a bandwidth of 7.4TB/s, showcasing improved energy efficiency over the TPU Trillium by twofold.

Google Cloud customers can opt for the Ironwood cluster in two sizes: 256 chips or a staggering 9,216 chips, reaching a full capacity of 42.5 ExaFLOPS. The TPU is exclusively available within Google’s cloud ecosystem, boasting superior performance-to-price ratio compared to its competitors. This is evidenced by the enhanced performance of Gemini LLM service over OpenAI’s GPT-4o and DeepSeek-R1, both relying on NVIDIA chips.

At the event, the popular vLLM library announced official support for TPUs, enabling seamless integration with PyTorch, JAX, vLLM, and Keras. Additionally, real-time Pathways developed by DeepMind for training AI models on large-scale clusters have been incorporated, allowing for scalable cluster adjustments based on usage.

Google plans to make TPU 7 Ironwood available for customer use within this year.

TLDR: Google introduced the TPU Ironwood chip at Google Cloud Next 2025, offering enhanced processing power and energy efficiency for large-scale AI applications.