Google Expands TPU 8 Series for Secure AI and Agentic Workloads
New TPU 8tx and 8i Max chips target confidential inference and long-context agent performance.
1Expanding the TPU 8 Family
Google has unveiled two new additions to its eighth-generation TPU lineup: the TPU 8tx for secure, hardware-isolated inference and the TPU 8i Max for long-context agentic tasks. These systems build upon the previously announced TPU 8t and 8i chips, aiming to provide more specialized hardware options for diverse enterprise AI requirements.
2Optimizing for Agentic Reasoning
The TPU 8i Max combines two 8i chips into a single package, effectively doubling compute, SRAM, and HBM bandwidth. This configuration addresses the memory wall challenge, where large models with million-token context windows require faster data access. It allows for more efficient handling of complex AI agents working across large software codebases.
3Hardware-Based Confidential Inference
Designed for regulated industries and government agencies, the TPU 8tx provides hardware-isolated inference. By utilizing 5th Gen AMD EPYC CPUs instead of Arm-based hosts, it ensures that sensitive data and model weights remain protected during processing. This setup enables organizations to fine-tune proprietary models without moving sensitive data between different clusters.
4Network and Latency Improvements
The 8tx adopts the Boardfly network architecture, which Google claims reduces latency by up to 2x compared to the 8t’s torus network. While torus networks are efficient for predictable training patterns, Boardfly optimizes inference by limiting the number of hops between distant chips to seven, even in large clusters of 1,024 chips.
5Software-Driven Memory Management
Google introduced TPU Sync and the Agent Compute Engine to maximize hardware efficiency. TPU Sync automatically tiers massive KV caches from HBM to host DRAM, reportedly cutting transfer latency by over 90%. Meanwhile, the Agent Compute Engine pauses idle sessions to save memory, further optimizing the price-performance ratio for complex agentic workloads.
6Competitive Performance Landscape
While official eighth-gen benchmarks are pending, previous-generation Ironwood TPUs demonstrated significant performance-per-dollar advantages over Nvidia’s B200 and B300 chips in specific FP8 tests. The new 8i Max aims to close the compute gap with Nvidia’s Rubin R200 and AMD’s MI455X by increasing memory bandwidth and compute density within existing rack footprints.
7Strategic Shift in Development
The rapid introduction of these specialized systems, just seven months after the initial TPU 8 launch, signals a shift toward modular, faster hardware iteration. By offering targeted solutions for security and long-context reasoning, Google seeks to capture market share from competitors like Nvidia and Intel, who are also developing confidential computing domains.
8Future Outlook and Availability
Both the TPU 8tx and 8i Max are currently listed as coming soon, with full technical specifications, pricing, and availability dates yet to be disclosed. Success will depend on Google’s ability to provide verifiable attestation for security and demonstrate consistent performance gains through third-party benchmarks once the hardware reaches customer hands.
Original article: Tom's Hardware — Google Expands TPU 8 Series for Secure AI and Agentic Workloads