Google Introduced a new AI accelerator chip called Ironwood

At its recent Cloud Next conference, Google introduced its latest innovation in AI processing hardware: the Ironwood TPU, the seventh-generation Tensor Processing Unit designed specifically for inference-based tasks. Unlike previous TPU iterations that were more general-purpose or training-focused, Ironwood is fine-tuned to efficiently run large-scale AI models, marking a pivotal step in Google’s advancement of AI infrastructure.

The Ironwood chip is expected to roll out later this year to Google Cloud customers and will be available in two scalable formats-a smaller 256-chip cluster and a massive 9,216-chip supercluster. This is Google’s “most powerful, capable, and energy-efficient TPU to date,” according to Google Cloud VP Amin Vahdat. It was developed from the ground up to satisfy the requirements of inferential AI at the business level.

This launch intensifies the ongoing battle in the AI accelerator landscape, where companies like Nvidia currently dominate, but rivals such as Amazon and Microsoft are quickly catching up with proprietary silicon of their own. Amazon’s Trainium, Inferentia, and Graviton processors are already accessible via AWS, while Microsoft is offering access to its custom-built Maia 100 chips through Azure.

Based on internal performance tests, Ironwood delivers an impressive 4,614 teraflops (TFLOPs) of peak computing performance. Each TPU is outfitted with 192GB of high-speed memory, capable of reaching bandwidths nearing 7.4 terabits per second. A standout feature of Ironwood is its upgraded SparseCore engine, purpose-built for handling data-intensive applications like personalized recommendations and advanced ranking systems-key components of many modern AI-driven services.

Ironwood’s architectural design minimizes data transfer and latency across its chips, translating into not only faster performance but also improved energy efficiency, a growing concern in large-scale computing. Google also intends to incorporate Ironwood into its modular AI Hypercomputer platform, enhancing flexibility and performance for cloud-based machine learning workloads.

“Ironwood symbolizes a major leap forward in inference computing,” Vahdat emphasized. “It offers superior computational performance, expanded memory resources, advanced networking, and the kind of dependability that enterprise AI workloads demand.”