
Google has unveiled its most powerful artificial intelligence (AI) processor, Ironwood, reports 3dnews.ru. This is the company's seventh-generation TPU (Tensor Processing Unit) and its first chip specifically optimized for inference — running already trained AI models.
Ironwood will be used in Google Cloud and is available in two configurations: servers with 256 chips and clusters with 9216 chips.
According to Google Cloud VP Amin Vahdat, Ironwood is the company’s most powerful, highest-performing, and energy-efficient TPU. It is designed to accelerate inference at cloud scale.
Ironwood delivers up to 4614 Tflops (around 4.6 quadrillion operations per second). A cluster of 9216 chips reaches a total performance of 42.5 exaflops.
Each chip has 192 GB of dedicated RAM with 7.4 Tbit/s bandwidth. It also includes a specialized SparseCore module tailored for workloads such as recommendation systems and advanced ranking.
Ironwood will be deployed in the AI Hypercomputer infrastructure within Google Cloud.