Alphabet Inc. on Wednesday introduced its latest artificial intelligence chip, named Ironwood, during its annual cloud conference, aiming to boost the speed and efficiency of AI applications.
The new chip is specifically designed for “inference” computing—the intensive data crunching required to deliver instant responses in applications like chatbots and generative AI tools.
The Ironwood processor brings together features previously split between Google’s earlier TPU designs, offering enhanced memory and cost-efficient performance for running large-scale AI models. It is engineered to operate in massive clusters of up to 9,216 chips, which the company says will significantly improve the delivery of AI-driven services.
Amid growing demand for faster and more energy-efficient AI infrastructure, Google said Ironwood doubles the performance per unit of energy compared to its Trillium chip announced last year. “The relative importance of inference is going up significantly,” said Amin Vahdat, Google vice president, highlighting a broader industry shift toward optimizing AI application performance.
Google’s tensor processing units (TPUs), which power many of its internal AI projects and its Gemini AI models, are available exclusively to its engineers or customers through its cloud service. The company’s long-term, multi-billion dollar investment into its custom chip ecosystem has positioned it as one of the few players capable of challenging Nvidia’s dominance in the AI chip market.
Google’s strategy has involved developing two separate TPU lines in the past: one tailored for training AI models from scratch, and another streamlined version designed to lower the costs of operating AI applications. Ironwood unifies these priorities, offering a powerful new option as companies worldwide race to deploy generative AI solutions.
