Google Unveils Two Eighth-Gen TPU Systems, Bringing Confidential Computing to TPU
Frontier AI models · Enterprise AI · Cybersecurity and privacy
Google announced two new eighth-generation TPU systems: TPU 8tx, featuring confidential computing with on-premises deployment options, and TPU 8i Max, designed for million-token context workloads and high-throughput AI agents. TPU 8i Max packages two 8i chips with 384GB of high-bandwidth memory. Google is also expanding framework support through the open-source TorchTPU project and native backends for inference engines including vLLM and SGLang.
To address enterprise demands for scaling AI agents, Google introduced two new eighth-generation TPU systems: TPU 8tx and TPU 8i Max. Google Cloud leadership noted that infrastructure demands have shifted toward agent scale and action-oriented systems requiring secure processing and long context retention. TPU 8tx builds on the TPU 8t chip combined with custom AMD host CPUs to provide hardware-level confidential computing. The data path from CPU across PCIe to the TPU is encrypted to protect prompts, weights, and enterprise data from exposure or operator snooping. The system supports Google Cloud and on-premises enterprise data center deployments, incorporating Broadfly network topology to reduce latency. TPU 8i Max integrates two 8i chips into a single module, providing 384GB of high-bandwidth memory (17.2TB/s bandwidth) and 768MB of SRAM cache to double throughput for concurrent agent workloads. To lower migration friction from GPUs, Google highlighted the TorchTPU initiative with Meta along with TPU backends for open-source inference engines vLLM and SGLang.
Key takeaways
- Google unveiled TPU 8tx with hardware-level confidential computing and support for select on-premises deployments.
- TPU 8i Max combines two chips to provide 384GB memory and 17.2TB/s bandwidth for long-context and multi-agent workloads.
- Google expanded software support via the TorchTPU project and backends for inference engines vLLM and SGLang.
Source：iThome Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.