Google advances its artificial intelligence technology through the launch of its seventh-generation Tensor Processing Unit (TPU), Ironwood. The custom chip represents a fundamental shift in Google’s hardware approach by delivering substantial improvements to serve the complex requirements of its top-tier Gemini models instead of just offering minor updates. Ironwood is engineered specifically to perform in simulated reasoning tasks that Google refers to as “thinking,” marking the dawn of a groundbreaking AI epoch.

Ironwood’s Design and Purpose

Ironwood achieves its capabilities through major improvements in both performance levels and architectural design advancements. Ironwood represents a considerable improvement in throughput performance and has been specifically built for operation in expansive liquid-cooled clustered environments when compared to previous TPU models. Clusters with up to 9,216 individual chips connect via an advanced Inter-Chip Interconnect (ICI) to enable efficient high-speed communication and data transfer. The scalable architecture enables Google’s research and development teams and external Google Cloud developers to utilize systems that vary from servers with 256 chips up to complete clusters with 9,216 chips.

Google’s Vision for AI

Google believes Ironwood’s advanced speed and power efficiency, combined with its larger memory capacity, will make a substantial impact on the AI ecosystem while enabling major advancements. Ironwood supplies a strong computational base to support advanced AI models, which will promote discoveries across multiple disciplines such as natural language processing and machine learning, as well as agentic AI development. The upcoming AI generation will function proactively by independently collecting data and analyzing information to take user-directed actions with limited guidance. Ironwood acts as a critical partner for Google while it pushes forward with AI advancements.

The Driving Force Behind Ironwood

Through the Ironwood project, Google manifests its belief that the synergy between cutting-edge AI models and specialized infrastructure is essential. Google describes Ironwood as an essential element of their strategy to boost inference speeds while extending AI model context windows and fully unleashing the capabilities of “agentic AI.” Google describes its new paradigm as the “age of inference,” where AI systems will take proactive actions to benefit users.

Ironwood’s Technical Specifications

The core specifications demonstrate Ironwood’s impressive computational capabilities. The peak inference computing performance of a fully configured Ironwood pod reaches 42.5 Exaflops. The peak throughput performance of each Ironwood chip stands at 4,614 TFLOPs which marks a significant advancement from earlier TPU models. Ironwood provides enhanced processing capabilities through its significantly advanced memory architecture. The new Ironwood chip features 192GB of high-bandwidth memory which represents a sixfold growth from the Trillium TPU. The memory bandwidth now stands at 7.2 Tbps which reflects a 4.5 times enhancement.

Benchmarking Ironwood

Google has published benchmarks that use FP8 precision as the primary indicator for measuring Ironwood’s performance. The company asserts that Ironwood “pods” deliver a 24-fold speed advantage over similar components of top supercomputers, yet this statement requires careful interpretation. Google recognizes that not all supercomputing systems support FP8 precision natively, which affects comparative assessments. The text does not present direct performance comparisons between Ironwood and Google’s TPU v6 (Trillium). According to Google, Ironwood delivers double the performance per watt compared to Trillium, which shows improved energy efficiency. The TPU v5p was succeeded by Ironwood, according to a Google spokesperson, while Trillium came after the TPU v5e. The top FP8 computing power of Trillium reached roughly 918 TFLOPS.