Google Introduces Ironwood, Its First Inference-Optimized AI Chip
Unveiled at Cloud Next 2025, Ironwood delivers unprecedented performance and efficiency for real-time AI applications, marking a shift in Google's AI strategy.
- Ironwood is Google's seventh-generation TPU, purpose-built for inference workloads, offering double the performance per watt compared to its predecessor, Trillium.
- The chip comes in two configurations: a 256-chip cluster and a 9,216-chip cluster, delivering up to 42.5 exaFLOPS of computing power for demanding AI tasks.
- Enhanced memory and inter-chip connectivity enable Ironwood to support larger AI models and reduce latency, with 192 GB of high-bandwidth memory per chip and 1.2 Tbps bidirectional bandwidth.
- Integrated into Google's AI Hypercomputer architecture, Ironwood is designed to power advanced models like Gemini 2.5, supporting scalable deployment for enterprise AI applications.
- Available exclusively through Google Cloud later in 2025, Ironwood positions Google to challenge Nvidia by addressing the cost and energy demands of AI inference at scale.
Wccftech 
Interesting Engineering 
The Register 
Maginative 
Google 
Ars Technica 
Investor's Business Daily 
ZDNET 
VentureBeat 
TechCrunch 