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

Google Cloud announced the seventh generation Tensor Processing Department (TPU) called Omasy On Wednesday, the company’s claims claims a special AI accelerator, when placed on a scale, the world’s fastest supercomputer is more than 24 times.

New chip, announced Google Cloud Next ’25The AI ​​chip represents an important Pivot in the length of the decade of Google. Previous generations are primarily designed to process both exercise and result work, the first goal is to prepare former AI models to make a railway, forecast or respond.

“The iron tree was established to support a generative AI and its great calculation and communication requirements,” he said. “It’s not only information, but to combine concepts and answers, we will actively get back to the AI ​​agents and learn to get information.”

Break calculation barriers: Ironwood’s 42.5 AI muscle 42.5 exaflops

Technical features Omasy Surprisingly. 9,216 chips per pod, iron tree calculation power 42.5 exaflopu – dwarf Captain‘S 1.7 exaflops, the fastest superkomputer in the world at the moment. Each individual provides a peak reports to the Ironwood Chip 4.614 Teraflops.

Ironwood also develops important memory and bandwidth. Each chip is 192GB high bandwidth memory (HBM), six times more TrilliumGoogle’s previous generation TPU announced last year. The memory bandwidth reaches 7.2 theretes every second to improve 4.5x over a trilium.

Perhaps most importantly, in one period of data centers restricted by electricity, Omasy Performs each home twice as compared to TrilliumAnd since 2018, Google’s first Cloud TPU is about 30 times more effective.

“When the existing power is one of the restrictions to convey AI capabilities, the customer provides significantly more power for a watt for workloads,” Vahdat said.

Model building ‘Thought Machines’ to ‘Machines’: Google’s result is why it is now important

Emphasis of more results than training represents an important infection point in the AI ​​schedule. For years, the industry has been increasing mass foundation models, primarily in parameter size and companies competing in training opportunities. Google’s Pivot, placement for the end of Google proposes to enter a new stage where the central stage of its placement effectiveness and reasoning opportunities.

This transition gives meaning. The training occurs once, but users occur repeatedly as daily billions of days as the interaction with AI systems. The EU’s economy is increasingly related to the expenses of inference, especially models, more complex and calculation intensify.

During the press conference, Vahdat, Google, in the last eight years, the report of the EU report consistent in the 2nd year – a 100 million factor in the report of 10,000 years. There is no amount of Moore Law Progress can provide this growth curve without specialized architecture as a railway tree.

What is especially noteworthy is the focus of “thinking models”, which performs more complex thinking tasks than recognition of a simple example. This can only simulate the future of Google in models that can disrupt problems, but also in models that can disrupt problems.

Gemini Thinking Engine: How Google uses the advanced apparatus of Next Gen Models

Google places the iron tree as the foundation of the most advanced AI models, including Twins 2.5“It describes the company as” Rightly built thinking skills. “

Google was also announced at the conference Twins 2.5 FlashA more efficient version of the flagship model “regulating the depth of justification based on urgent complexity. Gemini 2.5 Pro is designed for cases of intricate use and financial modeling, Gemini 2.5 Flash is placed for daily applications where sensitive is critical.

The company also displayed a complete set of media models for music for text-image, text-videos and music for a new announced text Liria. One demonstration showed how these vehicles can be used to create a complete presentation video for the concert.

Outside Silicone: Google’s comprehensive infrastructure strategy includes network and software

Omasy Google is just part of a broader AI infrastructure strategy. The company also announced Cloud waterGoogle’s planetary scale is a large networking service that provides access to special network infrastructure.

“Cloud Wan, a fully managed, vibrant and reliable enterprise network that provides 40% improved network performance, is a network support, but also has 40%.

Google is also expanding software offers for AI workload WayMachine learning time is developed by Google Deepmind. Roads on Google Cloud allow customers to save the model that serves in TPUs.

AI Economy: How does Google plan to win $ 12 billion in cloud business efficiency

This apparatus and program ads come in an important time for Google Cloud $ 12 billion in G4 2024 incomeDuring the last 30% annual year in the latest earnings report.

The economy of the EI placement is becoming a factor distinguished in cloud wars. Google faces strong competition Microsoft AzureThis is using it Openai partnership to a giant market position and Amazon Web Servicescontinues to expand it Training and Integrity Chip offers.

What separates Google’s approach is its vertical integration. Although opponents are partnerships with chip manufacturers or the beginnings of the googles, Google has been developing in TPU in TPU for more than a decade. This is unparalleled by Silicone from silicone to serve, unparalleled management with the AI ​​stack.

This technology will bring to enterprise customers, Google, Search, Gmail and YouTube, competitive advantages in the enterprise market. The strategy is clear: Google offers the same infrastructure that gives you power to anyone who wants to pay for it, on a scale of Google, on a scale.

Multiple Agent Ecosystem: Heart Plan for AI Systems that Google Working together

Outside the supply, Google has announced a vision for AI around many agent systems. The company has announced Agent Development Kit (ADK) This allows developers to build many AI agents able to work together.

Perhaps the most importantly, Google, different frameworks and AI agents built on different vendors and different vendors, “A2A) announced” A2A), which allows you to communicate with each other.

“2025 will be the year to answer the general questions of the Generative AI, to respond to the solution of complex problems with agent systems,” Vahdat was forecasted.

Google is partner with more than 50 industrial leaders Righteousness, Serviceenowand Leakto develop this inability standard.

Enterprise Reality Check: What does the power and efficiency of the railway tree for your AI strategy mean

These announcements for institutions in AI can significantly reduce the cost and complexity of advanced AI models. Improved effectiveness of Ironwood, more economical models, more economical, and the agent’s interaction protocol can help prevent the vendor lock.

The effect of the real world of these progress should not be evaluated. Many organizations relieved to place advanced AI models due to prohibited infrastructure costs and energy consumption. Google, according to his performance, was able to see a new wave of the EU adoption in the areas that are far outside.

Many agent approaches are extremely important with the complexity of placement between EU’s various systems and vendors. By standardizing how the AI ​​systems communicate, Google tries to break down the silos that restricts the impact of the limited EU enterprise.

During the press conference, Google stressed that more than 400 customer statements will be shared in the next ’25, and the effect of real work from AI innovations.

Silicone Weapon Race: Will Google’s special chips and open standards change the EU’s future?

As AI continues to advance, the strengthening of infrastructure will be increasingly critical. Google’s specialization like Google’s initiatives offers Google specializations, the company’s more distributed, more complex and more complex and deeper integrated.

The leading thinking models of the “Nobel Prize”, which is the “Twins 2.5” and the Nobel Prize, we did not expect to see what our developers and Google Cloud customers have in this year. “

The strategic results extend outside Google. When protecting property preferences in hardware, Google tries to act in subtle balancing. The company provides the prosperity of a wider ecosystem (with the Google Infrastructure), still compete.

The sooner the opponents respond to Google’s apparatus and the proposed agent will be important factors to eliminate industry around the standards of interaction. If history is any guide, we can expect Microsoft and Amazon to face their results to face optimization strategies, the most effective AI infrastructure stack builds a three-way race.