Ironwood: Google AI chip surpasses the best supercomputer in 24 times

During the Google Cloud Next 25 event, held this week, the technology giant officially revealed the Ironwoodits seventh generation of TPUS (Tensor Processing units).

The new accelerator comes with an unprecedented jump in performance and energy efficiency and was developed specifically to deal with inference tasks – the new frontier of artificial intelligence, according to the company itself.

Designed for the New Age of Inferential Age

Unlike its predecessors, which balanced processing between training and inference, the Ironwood was created from scratch to “thinking” models – Like LLMS (large -scale language models), moes (mixture of experts) and advanced reasoning tasks.

This is because the new category of AI not only responds to commands, but anticipates, interprets and generates insights proactivelystarting the so -called “Age of Inference”.

We are facing a paradigm shift. AI now needs not only to answer, but understand, anticipate and collaborate with the human being

Amin, VP of Google AI and Cloud Systems

24 times more powerful than supercomputer number 1

The most powerful configuration of Ironwood has 9,216 interconnected chipsdelivering impressive 42.5 Computational Power Exaflops – A number that Overcomes 24 times the performance of the most powerful supercomputer in the world, El Capitanwhich operates at 1.7 Exaflops.

Each individual chip has Maximum capacity of 4,614 TFLOPswhich demonstrates the exponential advance in relation to previous generations of TPUS.

Optimized architecture for extreme performance

Among the main architectural advances of Ironwood, stand out:

• Expanded HBM Memory: Each chip has 192 GB of high bandwidth memorysix times more than the previous generation, Trillium. This allows the processing of much larger models and data sets.
• Record Memory Band Width: they are 7.2 TB/S per chipproviding ultra -rare access to data, essential for memory intensive tasks.
• Interconnection between High Speed ​​Chips (ICI): com 1,2 bidirectional tbpsIronwood allows efficient synchronous communication between thousands of chips, ideal for distributed workloads.

It is noteworthy that the system has Advanced liquid cooling To ensure stable operation even under continuous and intense use, bending thermal efficiency compared to air -cooled systems.

Folded Energy Efficiency

Another highlight is the energy efficiency: Ironwood offers 2x more performance by Watt compared to Trillium.

This represents an important advance in a scenario where available energy limitation is one of the biggest obstacles to the evolution of AI on a large scale.

When compared to the company’s first cloud TPU, launched in 2018, Ironwood is almost 30 times more energetically efficient. However, Google has not yet stated which company behind the manufacture of semiconductors.

Comparative: Ironwood x Supercomputers (April 2025)

Note: 1 Exflop = 1 Quintal of Operations per second. Ironwood operates with optimized AI (FP8) floating point numbers, which provides more performance in specific workloads such as inference.

Thinking beyond traditional AI

With the Sparsecore dedicated acceleratorIronwood can also deal with grandes embeddings Used in recommendation and classification systems – essential features in sectors such as finance and scientific research, going beyond the traditional AI domains.

Thanks to Pathwaysdistributed execution system developed by Google Deepmind, developers can Connect tens or even hundreds of thousands of Ironwood tpusenabling the execution of tasks that were previously considered impossible.

What changes in practice? Why does it matter to you?

Although it seems too technical, the advances brought by Ironwood will directly impact the lives of millions of people. And in very visible ways:

• Models of AI much smarter and faster: Chatbots, virtual assistants and recommendation systems will be more accurate, natural and proactive.
• Automatic translations and real -time subtitles with almost perfectionincluding in complex contexts.
• Medical diagnoses for faster and more reliablebased on the analysis of gigantic images and historical clinical data.
• Accelerated science and research: Molecular, climate and engineering simulations will be performed in hours – which once took days or weeks.
• Automation of complex tasksas assisted programming, logistical or legal planning, will be brought to another level.
• Fastest and more creative generalenabling videos, voices, images and texts with professional quality in a matter of seconds.

In summary: AI will move from “good helper” to “real copilot” in the daily lives of people and companiesthanks to the unprecedented computational capacity that Ironwood delivers.

Also read:

Implications of the new technology

The arrival of Ironwood elevates the batten in relation to the standard of what is possible in high performance AI. Google Cloud positions itself as one of the main infrastructure players for artificial intelligence, breaking with Nvidia’s hegemony and entering the dispute with other giants such as Microsoft (with Maia 100) and Amazon (with the Graviton chips).

With the increasing adoption of models like Gemini 2.5 e AlphaFold -This last Nobel Prize winner-on TPUS, Ironwood is expected to be the engine of a new generation of scientific discoveries, innovative products and natural language advances.

Source: Google