Prime Highlights:
- Google introduces “Ironwood,” the company’s highest performance-per-watt TPU to date, created to scale AI inference workloads.
- Scales in clusters of up to 9,216 chips, Ironwood is optimized for real-time processing of heavy memory and bandwidth-intensive AI models.
Key Facts:
- Delivers 4,614 TFLOPs peak compute and includes 192GB RAM per chip.
- Tied together with SparseCore for recommendation and ranking workloads.
- To be included in Google Cloud’s AI Hypercomputer platform.
Important Backstory
From its Cloud Next 2025 conference, Google debuted Ironwood, its seventh-generation Tensor Processing Unit (TPU), custom-built for AI inference — processing pre-trained AI models in real time. The strategic hardware advancement is a watershed moment in Google’s quest to meet growing demands for scalable, efficient AI infrastructure.
Ironwood is built for scale performance and has two configurations that ship: one containing 256 chips and one massive cluster with support for up to 9,216 chips. The chip itself has a top computing capacity of up to 4,614 TFLOPs according to Google’s internal benchmarks. It also comes with 192GB of dedicated RAM per chip with 7.4 terabits per second bandwidth, which is essential in serving the advanced data streams in scale AI workloads.
One of the several features is the scalable core architecture of Ironwood, i.e., the SparseCore. The purpose-built processing engine is specifically optimized for recommendation system and complex data ranking workload — workload at the center of e-commerce, search, and social media algorithmics today. SparseCore optimizes on-chip data movement and removes latency, with performance gain and power savings.
This introduction also aligns with broader strategic alignment with Google’s AI Hypercomputer initiative, a modular compute cluster bringing AI infrastructure to users in more locations via Google Cloud. The inclusion of Ironwood in this environment will position customers to be in a position to be able to leverage its capability on cloud platforms in the near future, taking its commercial applicability even wider.
The technology is the result of fierce competition in AI chip technology. Even though Nvidia leads now, other tech industry titans are closing in fast using homegrown chips. Amazon already has Trainium, Inferentia, and Graviton processors, Microsoft just introduced Maia 100, and Google is now squarely set to not only match but even surpass inference-focused AI hardware with Ironwood.
