Google Ironwood: 24x Faster Than Supercomputers

Google’s Ironwood TPU: Redefining AI Inference

Google has unveiled its seventh-generation Tensor Processing Unit (TPU), Ironwood, a groundbreaking AI accelerator poised to reshape the landscape of artificial intelligence (AI) processing. In large-scale deployments, Ironwood’s computational prowess surpasses the world’s fastest supercomputer by over 24 times, marking a monumental leap forward in AI capabilities.

Announced at the Google Cloud Next ‘25 conference, this new chip represents a significant milestone in Google’s decade-long commitment to AI chip development. Unlike previous TPUs, which were versatile enough for both AI training and inference tasks, Ironwood is explicitly designed and optimized for inference workloads. This strategic shift underscores Google’s focus on maximizing AI deployment efficiency in real-world applications.

Amin Vahdat, Google’s Vice President and General Manager of Machine Learning, Systems, and Cloud AI, highlighted the significance of this transition, stating, “Ironwood is designed to support the next phase of generative AI and its immense compute and communication demands. This is what we call the ‘Inference Era,’ where AI agents will proactively retrieve and generate data to collaboratively deliver insights and answers, rather than merely processing data.”

Unprecedented Computing Power: 42.5 Exaflops

The technical specifications of Ironwood are nothing short of remarkable. When scaled to a pod consisting of 9,216 chips, it delivers an astounding 42.5 exaflops of AI compute. To provide context, this eclipses the current world’s fastest supercomputer, El Capitan, which operates at 1.7 exaflops. Each individual Ironwood chip boasts a peak compute capability of 4614 TFLOPs.

Beyond sheer processing power, Ironwood incorporates substantial enhancements in memory and bandwidth. Each chip is equipped with 192GB of high-bandwidth memory (HBM), representing a sixfold increase compared to the previous generation TPU, Trillium, released the previous year. Furthermore, the memory bandwidth per chip reaches 7.2 terabits/s, which is 4.5 times greater than that of Trillium.

• Compute Power: 42.5 exaflops (per pod of 9,216 chips)
• Peak Compute per Chip: 4614 TFLOPs
• Memory: 192GB HBM per chip
• Memory Bandwidth: 7.2 terabits/s per chip

In an era characterized by expanding data centers and growing concerns about power consumption, Ironwood demonstrates significant improvements in energy efficiency. It offers twice the performance per watt compared to Trillium and nearly 30 times that of the first TPU introduced in 2018. This leap in energy efficiency is crucial for the sustainable scaling of AI infrastructure.

The Inference Era: A Paradigm Shift

The optimization of Ironwood for inference workloads signals a pivotal turning point in the evolution of AI. In recent years, leading AI research labs have prioritized the development of increasingly large foundation models with ever-expanding parameter counts. However, Google’s focus on inference optimization suggests a strategic shift towards a new paradigm centered on deployment efficiency and real-world inference capabilities.

While model training remains essential for developing sophisticated AI, inference operations are far more frequent, occurring billions of times daily as AI technologies become more deeply integrated into everyday life. For businesses leveraging AI, the economics are intrinsically tied to inference costs as models become more complex. Efficient inference is therefore critical for cost-effective AI deployment at scale.

Google’s AI compute demand has grown tenfold in the past eight years, reaching an astounding 100 million. Without specialized architectures like Ironwood, it would be impossible to sustain this growth trajectory through traditional advancements in Moore’s Law alone. Specialized hardware is therefore essential for enabling future AI advancements.

Importantly, Google’s announcement emphasized a focus on “reasoning models” capable of executing complex inference tasks rather than simple pattern recognition. This suggests a belief that the future of AI lies not only in larger models but also in models capable of breaking down problems, engaging in multi-step reasoning, and emulating human-like thought processes. Such models require powerful inference capabilities and specialized hardware like Ironwood.

Powering Next-Generation Large Models: Gemini 2.5 and Beyond

Google is positioning Ironwood as the foundational infrastructure for its most advanced AI models, including its own Gemini 2.5, which boasts “native reasoning abilities.” This highlights the close integration between hardware and software in Google’s AI strategy.

The company also recently introduced Gemini 2.5 Flash, a smaller, more efficient version of its flagship model designed to “adjust reasoning depth based on the complexity of the prompt.” This model is geared towards everyday applications that require rapid response times, showcasing the need for specialized models tailored to different use cases.

Google further showcased its comprehensive suite of multimodal generation models, encompassing text-to-image, text-to-video, and its newly unveiled text-to-music capability, Lyria. A demonstration illustrated how these tools can be combined seamlessly to create a complete promotional video for a concert, highlighting the potential of multimodal AI for creative applications.

Ironwood is just one component of Google’s broader AI infrastructure strategy. The company also announced Cloud WAN, a managed wide area network service that provides enterprises with access to Google’s global-scale private network infrastructure. This enables faster and more reliable data transfer for AI workloads.

Google is also expanding its software offerings for AI workloads, including Pathways, a machine-learning runtime developed by Google DeepMind. Pathways now allows customers to scale model serving across hundreds of TPUs, demonstrating Google’s commitment to providing a comprehensive AI development and deployment platform.

Introducing A2A: Fostering an Ecosystem of Intelligent Agent Collaboration

Beyond hardware advancements, Google presented its vision for AI centered around multi-agent systems, unveiling a protocol to facilitate the development of intelligent agents: Agent-to-Agent (A2A). This protocol is designed to promote secure and standardized communication between different AI agents. A2A is envisioned as a key enabler of future AI applications.

Google believes that 2025 will mark a transformative year for AI, with the application of generative AI evolving from answering single questions to solving complex problems through intelligent agent systems. This transition requires robust communication and collaboration between agents.

The A2A protocol enables interoperability across platforms and frameworks, providing agents with a common “language” and secure communication channels. This protocol can be viewed as the network layer for intelligent agents, aiming to simplify agent collaboration in complex workflows. It empowers specialized AI agents to work together on tasks of varying complexity and duration, ultimately enhancing overall capabilities through collaboration. This allows for the creation of more complex and sophisticated AI systems.

Understanding A2A’s Functionality

Google provided a clear comparison between MCP and A2A protocols in its blog post:

• MCP (Model Context Protocol): Primarily focused on tool and resource management.
  • Connects agents to tools, APIs, and resources through structured input/output mechanisms.
  • Google ADK supports MCP tools, facilitating interoperability between various MCP servers and agents.
• A2A (Agent2Agent Protocol): Dedicated to facilitating collaboration between individual AI agents.
  • Enables dynamic multimodal communication between agents without the need for shared memory, resources, or tools.
  • Envisioned as an open standard driven and shaped by the community.
  • Practical examples can be observed using tools such as Google ADK, LangGraph, and Crew.AI.

Essentially, A2A and MCP serve complementary roles. MCP equips agents with access to essential tools, while A2A empowers these equipped agents to communicate and collaborate effectively with each other. This synergistic relationship between the protocols is crucial for enabling advanced AI applications.

The impressive list of partners announced by Google suggests that A2A is poised to receive similar attention and adoption as MCP. The initiative has already attracted over 50 companies to its initial collaboration cohort, including leading technology firms and top global consulting and system integration service providers. This widespread support underscores the importance of A2A for the future of AI.

Google emphasized the protocol’s openness, positioning it as the standard method for agents to collaborate, irrespective of the underlying technology frameworks or service providers involved. The company stated that it adhered to the following five key principles when designing the protocol with its partners:

1. Embrace Agent Capabilities: A2A focuses on enabling agents to collaborate in their natural, unstructured ways, even if they do not share memory, tools, and context. The ultimate goal is to enable genuine multi-agent scenarios without limiting agents to mere “tools.” This allows for more flexible and adaptable collaboration.
2. Build on Existing Standards: The protocol leverages existing popular standards, including HTTP, SSE, and JSON-RPC, making it easier for enterprises to integrate with their existing IT infrastructure and stacks. This reduces the barrier to adoption.
3. Secure by Default: A2A is designed to support enterprise-grade authentication and authorization mechanisms, comparable to OpenAPI’s authentication schemes at launch. This ensures that communication between agents is secure and protected.
4. Support Long-Running Tasks: A2A is designed with flexibility in mind, enabling support for a wide range of scenarios, from quick tasks to in-depth research that may take hours or even days (especially when humans are involved). Throughout the process, A2A can provide users with real-time feedback, notifications, and status updates. This is crucial for complex and time-consuming tasks.
5. Modality Agnostic: The world of agents is not limited to text. Recognizing this, A2A is designed to support various modalities, including audio and video streams, enabling richer and more natural communication between agents.

Practical Application: Streamlined Hiring Process via A2A

Google provided a compelling example illustrating how A2A can significantly streamline the hiring process.

Within a unified interface like Agentspace, a hiring manager can delegate the task of finding suitable candidates based on specific job requirements to an intelligent agent. This agent can then interact with specialized agents in specific fields to efficiently complete candidate sourcing. Furthermore, the user can instruct the agent to schedule interviews and enable other specialized agents to assist with background checks, thereby enabling a fully automated, cross-system collaborative hiring process. This showcases the potential of A2A to revolutionize business workflows.

Google Embraces MCP: Joining the Model Context Protocol Ecosystem

In a significant move towards greater collaboration and standardization within the AI community, Google is also embracing MCP. Just weeks after OpenAI announced its adoption of Anthropic’s Model Context Protocol (MCP), Google followed suit and joined the initiative.

Google DeepMind CEO Demis Hassabis announced on X that Google will add support for MCP to its Gemini models and SDKs, although a specific timeline was not explicitly provided. This demonstrates Google’s commitment to open standards and interoperability.

Hassabis stated, “MCP is an excellent protocol that is rapidly becoming the open standard for the age of AI agents. We look forward to working with the MCP team and other partners in the industry to advance the development of this technology.”

Since its release in November 2024, MCP has quickly gained popularity and widespread attention, emerging as a simple and standardized way to connect language models with tools and data. Its ease of use and widespread applicability have made it a valuable asset for AI developers.

MCP enables AI models to seamlessly access data from various data sources, including enterprise tools and software, to complete complex tasks and to access content libraries and application development environments. The protocol allows developers to establish bidirectional connections between data sources and AI-driven applications, such as chatbots, enhancing their capabilities and functionality.

Developers can expose data interfaces through MCP servers and build MCP clients (such as applications and workflows) to connect to these servers. Since Anthropic open-sourced MCP, several companies have integrated MCP support into their platforms, fostering a vibrant and collaborative ecosystem.

Key Concepts: Ironwood, A2A, and MCP

To fully grasp the impact and significance of Google’s recent announcements, it’s essential to delve deeper into the core components: Ironwood, A2A, and MCP.

Ironwood: A Closer Look at the Inference Era

The shift from a primary focus on training models to optimizing for inference is a crucial evolution in the AI landscape. Training involves feeding massive amounts of data to a model to teach it to recognize patterns and make predictions. Inference, conversely, is the process of applying a trained model to make predictions on new, unseen data.

While training is a resource-intensive, often one-time (or infrequent) event, inference happens continuously and at scale in real-world applications. Consider the following examples:

• Chatbots: Responding to user queries in real-time requires efficient inference.
• Recommendation Systems: Suggesting products or content based on user preferences relies on rapid inference.
• Fraud Detection: Identifying fraudulent transactions as they occur necessitates real-time inference.
• Image Recognition: Analyzing images to identify objects, people, or scenes demands efficient inference processing.

These diverse applications require rapid, efficient inference to deliver a seamless and responsive user experience. Ironwood is meticulously designed and optimized specifically to excel at these computationally intensive tasks.

Key Advantages of Ironwood for Inference:

• High Throughput: The massive compute power (42.5 exaflops) empowers Ironwood to handle a massive volume of concurrent inference requests simultaneously.
• Low Latency: The high-bandwidth memory (HBM) and carefully optimized architecture minimize the time it takes to process each individual inference request, resulting in lower latency.
• Energy Efficiency: The improved performance per watt significantly reduces the operational costs associated with running large-scale inference deployments. This is crucial for sustainable AI deployment.

By prioritizing and optimizing for inference, Google is enabling businesses to deploy AI-powered applications more efficiently and cost-effectively, making AI moreaccessible and practical for a wider range of use cases.

A2A: The Foundation for Collaborative AI Systems

The Agent-to-Agent (A2A) protocol represents a significant step towards the creation of more sophisticated and collaborative AI systems. In a multi-agent system, multiple AI agents work together in a coordinated manner to solve a complex problem. Each agent may possess its own specialized skills and knowledge, and they communicate and coordinate with each other to achieve a common, overarching goal.

Consider a scenario involving automated customer support as a practical example:

• Agent 1: Understands the customer’s initial query and identifies the underlying issue.
• Agent 2: Accesses a knowledge base to find relevant information to address the customer’s issue.
• Agent 3: Schedules a follow-up appointment with a human agent if necessary, escalating complex cases appropriately.

These agents need to be able to communicate and seamlessly share information to provide a cohesive and positive customer experience. A2A provides the essential framework to enable this type of sophisticated collaboration.

Key Benefits of A2A:

• Interoperability: Enables agents developed on different platforms and frameworks to seamlessly communicate with each other, fostering a diverse and interconnected AI ecosystem.
• Standardization: Provides a common “language” and a well-defined set of protocols for agent communication, ensuring consistency and compatibility.
• Security: Ensures secure communication between agents, protecting sensitive data and maintaining data integrity throughout the collaboration process.
• Flexibility: Supports a wide range of communication modalities, including text, audio, and video, enabling richer and more natural communication between agents.

By actively fostering collaboration between AI agents, A2A enables the development of more powerful, versatile, and adaptable AI systems capable of tackling complex challenges across various domains.

MCP: Bridging the Gap Between AI and Real-World Data

The Model Context Protocol (MCP) directly addresses the crucial challenge of connecting AI models to the vast amounts of data required to perform their tasks effectively. AI models critically need access to real-time data from various sources, such as databases, APIs, and cloud services, to make accurate predictions and informed decisions.

MCP provides a standardized and efficient way for AI models to access and seamlessly interact with these diverse data sources. It defines a clear set of protocols for:

• Data Discovery: Facilitating the identification of the available data sources that are relevant to the AI model’s task.
• Data Access: Streamlining the retrieval of data from the identified data sources in a consistent and reliable manner.
• Data Transformation: Converting the data into a format that the AI model can readily understand and process, ensuring compatibility and efficient utilization.

By providing a standardized interface for data access, MCP simplifies the overall process of integrating AI models with real-world data, enabling them to perform more effectively and deliver greater value.

Key Advantages of MCP:

• Simplified Integration: Makes it significantly easier to connect AI models to a wide range of data sources, reducing the complexity of AI development and deployment.
• Standardization: Provides a common and consistent set of protocols for data access, ensuring interoperability and reducing integration efforts.
• Increased Efficiency: Reduces the time and effort required to access and transform data, accelerating the development and deployment of AI applications.
• Improved Accuracy: Enables AI models to access the most up-to-date and relevant information, leading to more accurate predictions and improved decision-making capabilities.

By connecting AI models to the data they need in an efficient and standardized manner, MCP empowers them to perform more effectively, deliver greater value, and unlock new possibilities across various industries and applications.