Google’s new Ironwood chip is 24x more powerful than the world’s fastest supercomputer

Google Cloud unveiled the seventh generation tensioner unit (TPU)Ironwood, Wednesday. This dedicated acceleration of artificial intelligence claims the company, It provides more than 24 times the computing power of the fastest supercar in the world when it is widely spread.

The new slide, announced in Google Cloud Next’s 25, is a great axis in the AI ​​chips strategy that lasted for a decade of Google. Although the previous generations of TPUS are mainly designed for both the burdens of training and inference, Ironwood is the first specially designed to infer-the process of publishing artificial intelligence models trained to create predictions or generation of responses.

“Ironwood is designed to support this next stage of artificial intelligence and its computer requirements and enormous communication,” said Amin Fahdat, Vice President of Google, ML, Systems and Cloud AI, at a virtual press conference before the event. “This is what we call” the age of reasoning “, where artificial intelligence agents will recover and create data in a proactive way to provide cooperative visions and answers, not just data.”

Shattering mathematical checkpoint

Technical specifications for Ironwood striking. When it is limited to 9,216 chips for each pod, Ironwood 42.5 Exaflops of Computing power – El Capitan’s 1.7 Exaflops, is currently the fastest superconduction in the world. Each chip of Ironwood is an individual peak account of 4,614 teraflops.

Ionwood is also characterized by large memory and domain width. Each slide comes with the 192 GB high -frequency domain memory, which is more than six times more than Trillium, according to the previous generation TPU of Google last year. The memory domain width of the memory reaches 7.2 TERABITS per second for each chip, which is an improvement of 4.5X on Trillium.

Perhaps more importantly, in the era of energy -restricted data centers, Ironwood offers twice as much as watt compared to trillium, which is nearly 30 times more efficient in energy than the first cloud of Google than 2018.

“While the available energy is one of the restrictions imposed on providing the capabilities of artificial intelligence, we offer a much greater capacity per watt of customer work burdens,” Vahdat explained.

From building models to “Thinking Machines”: Why is it important to focus on Google’s inference now

The focus on inference instead of training is a great turning point in the Timet of Amnesty International. The industry has been installed on the construction of increasingly huge basic models for years, with companies competing mainly for the size of the teacher and training capabilities. The Google axis indicates improving the reasoning that we are entering a new stage as it takes the efficiency of publishing and the center of thinking.

This transition is logical. Training occurs once, but inferences occur billions of times daily as users interact with artificial intelligence systems. Artificial intelligence economics are increasingly associated with the costs of inference, especially since models grow more complicated and intense.

During the press conference, Vahdat revealed that Google has noticed a 10x increase on an annual basis in demand at the expense of artificial intelligence over the past eight years-an amazing factor of 100 million. None of the progress of the Moore law can satisfy this growth curve without specialized structures such as Ironwood.

What is particularly noticeable is to focus on “thinking models” that perform complex thinking tasks instead of identifying simple patterns. This indicates that Google sees the future of artificial intelligence not only in the larger models, but in models that can break problems, the reason through multiple steps and simulate human -like thinking.

GIMINI Thinking Engine: How the next Google models benefit from Google’s advanced devices

Google puts Ionwood as the basis for the most advanced artificial intelligence models, including Gemini 2.5, which the company describes as “integrated thinking possibilities”.

At the conference, Google also announced Gueini 2.5 Flash, a more cost -effective version of its main style “adjusts the depth of thinking based on the complexity of the claim.” Although Gemini 2.5 Pro is designed for complex use cases such as drug detection and financial modeling, Guemini 2.5 Flash is in a daily applications position where response is very important.

The company also showed a full range of obstetrics models, including text to a picture, a text from text to video, and a newly announced Music text capacity called Lyria. A demonstration showed how these tools can be used together to create a full promotional video of a concert.

Beyond Silicon: The comprehensive infrastructure strategy of Google includes network and programs

Ironwood is just one part of the broader infrastructure strategy than Google. The company also announced the Cloud Wan, a large -scale network service that allows companies to access the Google network infrastructure across the planet.

“Cloud Wan is a poor column of networks of fully managed companies, advanced and safe, providing up to 40 % on the network, while reducing the total cost of ownership by 40 % itself,” said Vahdat.

Google also expands its software shows for the burdens of artificial intelligence, including tracks, and the time of the automatic learning time developed by Google DeepMind. Google Cloud tracks allow customers to expand the model that works across hundreds of TPUS.

Artificial Intelligence Economics: How to plan with $ 12 billion clouds to win the efficiency war

These devices and programs are decisively for Google Cloud, which reached $ 12 billion in the revenue of the fourth quarter 2024, an increase of 30 % on an annual basis, in the latest profit report.

The economies of spreading artificial intelligence have become a growing factor in cloud wars. Google faces an intense competition from Microsoft Azure, which has benefited from its partnership Openai on a huge market site, Amazon Web Services services, which continue to expand training chips and trainees.

What separates Google’s approach is its vertical integration. While competitors have partnerships with drilling companies or startups acquired, Google is developing TPUS at home for more than a decade. This gives the company unparalleled on artificial intelligence, from silicone to programs to services.

By bringing this technology to institution agents, Google bets that its hard -class experience to search for Search, Gmail and YouTube will translate into competitive advantages in the institution market. The strategy is clear: He provided the same infrastructure that operates the Google International Amnesty International, on a large scale, to anyone who wants to pay for it.

Multi -deer ecosystem: The bold Google Plan for AI operating together

Outside of devices, Google has set a vision of AI centered on multi -agent systems. The company has announced the ADK development group (ADK), which allows developers to build systems where Amnesty International customers can work together.

Perhaps the most important thing, Google has announced “the agent of the agent’s interconnection to the agent” (A2A) that enables artificial intelligence agents integrated on different frameworks and different sellers to communicate with each other.

“The year 2025 will be transitional years, as artificial intelligence moves from answering one questions to solving complex problems through mummified systems,” Fahd predicted.

Google cooperates with more than 50 industry leaders, including Salesforce, Servicenow and SAP, to enhance this inter -operating standard.

Check the reality of the institution: What means the strength and efficiency of Ionwood for your artificial intelligence strategy

For institutions that spread artificial intelligence, these ads can significantly reduce the cost and complexity of advanced artificial intelligence models. Ironwood’s improved efficiency can make advanced thinking models more economical, while the agent’s interconnection protocol can help companies avoid locking the seller.

The effect of the real world of these developments should not be reduced. Many organizations were reluctant to spread advanced AI models due to the costs of high infrastructure and energy consumption. If Google is able to fulfill its performance promises for every wave, we can see a new wave of adopting artificial intelligence in the industries that have been on the margin.

The multi -agent approach is equally important for institutions overwhelmed by the complexity of spreading artificial intelligence through different systems and sellers. By uniting how artificial intelligence systems continue, Google tries to break silos that limit the foundation of the institution from artificial intelligence.

During the press conference, Google confirmed that more than 400 client stories will be shared next ’25, as it has a real impact of AI’s business innovations.

ARMS Silicon Race: Will you return the custom Google chip and open standards AI?

As artificial intelligence progresses, its infrastructure will be increasingly decisive. Google’s investments in specialized devices such as IONWOOD and overlapping operating initiatives indicate that the company determines itself for a future as artificial intelligence becomes more distributed, more complicated and deeper into commercial operations.

“The leading thinking models such as Gemini 2.5 and Alphafold, Nobel Prize winner, are all managed on TPUS today,” VAHDAT indicated. “With Ironwood, we can’t wait to find out what Amnesty International breakthroughs raised by our developers and Google Cloud developers when it becomes available later this year.”

Strategic effects go beyond Google’s business. By clicking on the open standards in the agent’s connection while maintaining the advantages of ownership in the devices, Google tries an accurate budget law. The company wants to flourish the broader ecosystem (with Google under it) while maintaining competitive distinction.

In the coming months, the main factors will include the speedy response of competitors to the developments of Google devices and whether the industry gather around the proposed inter -operating standards of the agent. If the date is any evidence, we can expect Microsoft and Amazon to deal with their inference improvements, which may create a three -way race to build the most efficient artificial infrastructure staple.