Google Unveils TPU 8tx and 8i Max for the Agentic Era

Google expands its eighth-generation TPU portfolio with TPU 8tx and TPU 8i Max, marking four new systems in a single year. These additions target secure, confidential inference and massive context windows, respectively, to support complex agentic workloads. The announcement highlights specialized hardware for security, density, and reasoning, alongside new software updates to improve accessibility for developers.

Audio reading is not available in this browser
Google Unveils TPU 8tx and 8i Max for the Agentic Era

Tags

Quick summary

Google expands its eighth-generation TPU portfolio with TPU 8tx and TPU 8i Max, marking four new systems in a single year. These additions target secure, confidential inference and massive context windows, respectively, to support complex agentic workloads. The announcement highlights specialized hardware for security, density, and reasoning, alongside new software updates to improve accessibility for developers.

Google Unveils TPU 8tx and 8i Max for the Agentic Era

Google has expanded its eighth-generation TPU portfolio with the introduction of two new systems: TPU 8tx and TPU 8i Max. According to Google AI, this announcement marks the addition of four new TPU systems in a single year, signaling a strategic shift toward specialized hardware designed for the "agentic era." While the previous generation introduced TPU 8t for large-scale training and TPU 8i for real-time reasoning, these new additions address specific challenges arising as enterprises deploy autonomous agents: hardware-level data security and massive context windows.

Specializing for Secure, High-Context AI

The primary driver for these new systems is the evolution of AI workloads from simple inference to complex, multi-step agent operations. Google notes that as computing moves from human to agent scale, businesses require accelerators capable of reasoning across massive context windows and handling sensitive data without compromising security.

TPU 8tx addresses the critical need for confidential inference. In traditional computing, data is encrypted at rest or in transit but must be decrypted in memory during processing, creating a vulnerability window. TPU 8tx, built on the same silicon as TPU 8t and paired with 5th Generation AMD EPYC CPU hosts, integrates security directly into the hardware layer. This architecture keeps sensitive data and model weights protected and isolated even while in active use.

Google highlights that TPU 8tx is optimized for regulated enterprises and governments running frontier or proprietary models alongside their most sensitive data. The system supports private fine-tuning, model distillation, and reinforcement learning for agents without requiring the migration of sensitive data between separate clusters. Furthermore, by adopting the same low-latency Boardfly network used for real-time reasoning, TPU 8tx delivers up to 2x lower latency compared to TPU 8t with a near-zero performance penalty. This ensures that multi-step agents can move efficiently while maintaining strict security protocols.

Breaking the Memory Wall with TPU 8i Max

While security is paramount for some workloads, others face the "memory wall." When dozens of AI agents collaborate across vast software codebases or analyze hours of video simultaneously, memory bandwidth often becomes the bottleneck, causing processors to stall.

TPU 8i Max is engineered specifically to solve this problem. The system combines two low-latency TPU 8i chips onto a single package. This dual-chip configuration scales compute and memory in lockstep, delivering double the compute, double the SRAM, and double the HBM bandwidth of the standard TPU 8i.

According to Google, this design ensures processors remain equipped with essential data, preventing large workloads from stalling. Additionally, because TPU 8i Max fits twice the silicon within the existing rack footprint, data centers can deploy significantly more computing power per square foot and per watt. This high compute density is crucial for handling the massive context windows required by modern agentic workflows.

A Modular Approach to System Balance

The introduction of TPU 8tx and TPU 8i Max reflects a foundational architectural decision made over a decade ago: building the computer around the chip. Google emphasizes that while designing a new chip involves a multi-year manufacturing lifecycle bound by the laws of physics, designing a new system allows for rapid adjustments to packaging, high-speed memory, optical networking, and security.

This modularity allows Google to deliver specialized supercomputers in a matter of months rather than years. The expanded eighth-generation portfolio now includes four distinct systems, each targeting a specific stage of the AI lifecycle:

  • TPU 8t: Designed for scale-up training with maximum "goodput" (useful intelligence delivered per watt).
  • TPU 8i: The reasoning engine designed to break the memory wall by cutting inter-chip communication delays.
  • TPU 8tx: Optimized for secure and confidential inference and small-scale training.
  • TPU 8i Max: Built for massive context windows and the highest compute density per rack.

This diversity allows customers to select purpose-built accelerators for every stage of the agentic lifecycle, from training foundation models to executing autonomous agents in production.

Opening Up TPUs to Every Developer

To complement the new hardware, Google is simultaneously working to remove barriers to adoption, aiming to make TPUs as easy to use as general-purpose hardware. The company has open-sourced TorchTPU, allowing PyTorch code and Hugging Face models to run natively on TPUs alongside out-of-the-box backends for inference engines like vLLM and SGLang.

Furthermore, a new open-source TPU agent suite (currently in preview) integrates directly with coding tools like Claude Code, Cursor, and Google Antigravity. This suite automates TPU onboarding, diagnoses bottlenecks, and autotunes custom kernels using plain language. For production deployments, the new open-source XLA Serving Engine enables teams to train models in Python and serve them directly in a low-latency C++ runtime across TPUs, GPUs, and CPUs.

Google also highlights improvements in production efficiency. The Agent Compute Engine software automatically suspends and resumes idle sessions for stateful AI agents, cutting operational infrastructure costs during pauses caused by database lookups or external tool interactions. Additionally, the open-sourcing of TPU Sync, a model-state layer for TPU clusters, unifies memory, storage, and networking in software. This layer is designed to handle the memory demands of agents reasoning across million-token contexts and continuous reinforcement learning loops.

Conclusion

The launch of TPU 8tx and TPU 8i Max represents a significant expansion of Google's infrastructure capabilities for the agentic era. By addressing the dual challenges of secure confidential inference and massive context window management, Google aims to provide the specialized hardware necessary for the next generation of AI applications. Coupled with recent software updates that lower the barrier to entry for developers, these systems position the eighth-generation TPU portfolio to support a wide range of companies and use cases, from frontier labs to regulated enterprises.

Source: https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx

Sources