Overview

• Ironwood, Google's seventh-generation TPU, is purpose-built for inference tasks, marking a strategic pivot from training-centric designs.
• The chip delivers 4,614 TFLOPs of peak performance, 192GB of memory, and 7.4 Tbps of bandwidth, doubling energy efficiency compared to its predecessor, Trillium.
• Ironwood features an enhanced SparseCore for handling complex recommendation and ranking tasks, making it ideal for data-intensive workloads.
• It will integrate into Google Cloud’s AI Hypercomputer, supporting configurations of up to 9,216 chips for scalable, high-performance inference workloads.
• This launch highlights Google's focus on reducing the rising operational costs of inference computing as AI applications become more widespread and resource-intensive.