During Cloud Next 2025, Google revealed its Ironwood – the seventh generation Tensor Processing Unit (TPU) – and a considerable advancement in the power of AI computing. Ironwood, according Google, represents a marked change in the development of AI and the infrastructure that supports its progress. It is being touted as the most powerful, scalable and energy-efficient AI chip ironwood has ever produced.

Ironwood is also the first TPU designed primarily to perform inference – the step during which AI systems respond, reason, and make decisions in real-time. This gives it the ability to use the vast amount of computational resources offered by large language models, mixture-of-expert (MoE) architectures, and sophisticated reasoning tasks with unprecedented speed and efficiency. Like every Ironwood chip, its performance exceeds expectations at 4,614 TFLOPs of peak compute. Google Cloud clients can then further scale their usage to an astonishing 9,216 chip pods. This astonishing number translates to over 42.5 exaflops of raw compute at full scale. Over 24 more than the current fastest supercomputer El capitan.

Per chip, it now has 192GB of ultra-fast memory, significantly increasing from before, while Google claims Ironwood doubles the performance-per-watt TPUs’ output last year. Due to the chip design limitations, they cannot move data and latency, low voltage Ironwood was developed with “thinking” AI models in mind.

Google provided additional information: “Our advanced liquid cooling solutions, and designed chips strive to ensure imminent, constant, aggressive AI workload will not exceed twice performance measured on air cooled systems. Ironwood is nearly 30x more power efficient than our first Cloud TPU in 2018, which makes thoroughly outweighing expenditure  on investment.”

With greater computational power, memory capacity, and ICI networking Ironwood, Google said: “marking the start of the age of inference” Technology level where AI is able to answer astutely without human intervention. Ironwood distinguishes itself as a one of a kind leap forward but not limited to reliability and immense computation power endurance on the era of inference, adding Google.

Read More: Google Unveils Ironwood: Its Most Powerful TPU Yet, Ushering in the ‘Age of Inference’