Google adds “8tx” and “8i Max” to eighth-generation TPU lineup; 8tx uses AMD CPU

On November 9, 2026, Google announced two new eighth-generation TPU systems, the TPU 8tx and TPU 8i Max. The 8tx pairs the same chip as the TPU 8t with an AMD fifth-generation EPYC processor and supports AI processing involving sensitive information. According to GIGAZINE, the 8i Max combines two chips to double processing performance.

For AI processing involving sensitive data, the TPU 8tx is designed to protect not only the TPU but also the host CPU and connected accelerators. AMD says the fifth-generation EPYC’s SEV feature supports that protection. The announcement concerns computing environments for companies and government agencies using sensitive information for additional training and inference.

The focus of the TPU 8tx is not the introduction of new accelerator processing units, but the integration of the existing TPU 8t, a host CPU, networking and security features into a computing environment for handling sensitive data. According to GIGAZINE’s report on Google’s announcement, the 8tx uses the same chip as the TPU 8t while Boardfly reduces latency by up to half. AMD’s fifth-generation EPYC handles processing orchestration and data preparation, and AMD says SEV supports extending protection to connected accelerators. Differentiation, therefore, appears to be based not only on the performance of the TPU itself, but on the system as a whole, including the path data takes from the host to the accelerator and protection during computation. Meanwhile, the 8i Max was reported to combine two chips to double processing performance. While the 8tx emphasizes confidentiality and low latency, the 8i Max is designed to improve performance in workloads that reference large amounts of information. This suggests that the eighth-generation product lineup is being divided by use case rather than by a single performance metric. However, the available information does not specify the benchmark used for the 8i Max’s performance comparison, measurement conditions or memory capacity, so the figure of 2x alone is not enough to compare actual throughput or usage costs. In a related article, we covered Google DeepMind’s announcement of the virtual-world generation AI “Genie 4” in Google DeepMind announces “Genie 4,” an AI for generating virtual worlds that also supports audio and conversation. That announcement concerned generating virtual environments for applications such as robot training, while the latest TPU system announcement concerns computing infrastructure. The two cannot be treated as evidence of the same implementation. The systems also have different roles from those discussed in Maxinsights raises $30 million to expand data infrastructure for Physical AI, which covered the collection and preparation of real-world data. Supplying training data, providing computing infrastructure and evaluating robots in the field are separate stages; the TPU announcement alone does not establish real-world robot performance or outcomes from data use. Points to watch include the specific conditions for the protection provided by the 8tx and the workloads it supports, the benchmark for the 8i Max, and the measurement conditions for latency, performance and power consumption in actual deployments. AMD’s announcement describes the role of the CPU and security features, but the information available does not allow an assessment of performance impacts in actual use or the conditions under which protection applies in individual environments. The products’ use-case-specific designs should be tracked separately from operational results and performance figures subject to specific conditions.

The TPU 8tx includes hardware features that protect and isolate data and AI model weights during processing. Google says its low-latency networking technology, Boardfly, reduces processing latency by up to half compared with the TPU 8t. It also supports small-scale training workloads, including additional training, distillation and reinforcement learning using enterprise data. According to AMD, the fifth-generation EPYC serves as the TPU 8tx’s host CPU, handling processing orchestration, data preparation and high-speed communication with the TPU. The system combines AMD’s Secure Encrypted Virtualization (SEV) feature with the TPU to extend confidential computing protection from the CPU to connected accelerators. AMD calls the TPU 8tx the first cloud TPU system to support hardware-enforced confidential computing.