Google advances its artificial intelligence efforts through the launch of its seventh-generation Tensor Processing Unit (TPU), which carries the name Ironwood. The introduction of this custom-designed chip shows Google making a substantial leap in hardware development by addressing complex requirements for its most sophisticated Gemini models instead of just implementing incremental updates.

The Ironwood chip stands out for its superior performance in simulated reasoning tasks that Google equates with “thinking” and represents Google’s step towards a new AI age. The performance and architectural design of Ironwood have been substantially enhanced to deliver its capabilities. Ironwood delivers a major throughput improvement over previous TPUs while being built for use in extensive liquid-cooled cluster systems. Google’s approach to hardware development supporting its AI goals has undergone a fundamental transformation.

The latest Inter-Chip Interconnect (ICI) enables high-speed data exchange between clusters with up to 9,216 individual chips. The interconnect technology enables efficient AI workload scaling, which allows massive computational power from these clusters to address complex problems.

Google Cloud provides external developers and its internal research and development teams with a scalable architecture that accommodates configurations from servers with 256 chips to full clusters with 9,216 chips. The fully configured Ironwood pod reaches an extraordinary performance level by delivering 42.5 Exaflops for inference computing. The peak performance of every single Ironwood chip reaches 4,614 TFLOPs, which marks an important advancement beyond earlier TPU generations.

Ironwood includes a highly improved memory architecture to back its enhanced processing capabilities. The HBM capacity for each chip reaches 192GB, which represents six times the memory of the Trillium TPU.

The expanded on-chip memory capacity plays a critical role in executing modern AI workloads because it enables the processing of large models and datasets while minimizing data transfer operations and boosting performance.

The memory bandwidth has improved substantially to reach 7.2 Tbps, which represents a 4.5 times increase. The processing units receive data at an optimal rate due to increased bandwidth, which ensures their full utilization for maximum efficiency.

Google predicts that Ironwood’s improved performance metrics will revolutionize its AI infrastructure, enabling major advancements. The robust computational foundation provided by Ironwood will enable increasingly sophisticated AI models to achieve breakthroughs across multiple domains, such as natural language processing and agentic AI development alongside machine learning. The upcoming AI generation will operate proactively by independently collecting information and executing actions based on minimal user instructions.

Ironwood serves as an essential partner in Google’s ongoing pursuit of AI advancement. Ironwood offers more than sheer processing strength because it serves as a platform for innovative AI applications and experiences.

Google supplied benchmarks for Ironwood performance by using FP8 precision as the main evaluation standard. The company states that Ironwood “pods” deliver a 24 times faster performance benchmark than equivalent parts of global top supercomputers, though this claim requires nuanced analysis. Google recognizes that the absence of native FP8 precision support in certain supercomputing systems affects their performance comparison.

The evaluation did not incorporate direct performance comparisons with Google’s TPU v6, known as Trillium. According to Google, Ironwood delivers double the performance per watt compared to Trillium, showing improved energy efficiency.

According to a Google representative, Ironwood replaces the TPU v5p, and Trillium is the following generation after TPU v5e. The maximum FP8 performance benchmark for Trillium reached about 918 TFLOPS. The design of modern AI hardware must prioritize energy efficiency due to the growing power requirements of these systems.