Google on November 9 announced two additions to its eighth-generation TPU lineup: the TPU 8tx, for confidential inference, and the TPU 8i Max, for large context windows.
Google said the systems respond to two customer demands: accelerators that can reason across very large context windows as computing moves "from human to agent scale," and, from "some highly regulated customers," specialized encryption capabilities for unique workloads. According to Google's post1, the additions mean four TPU systems in a single year, which Google described as a first.
Systems, not new chips
Google's post stresses that the two products are systems rather than new chips. It defines a system as the compute, networking and storage built around a chip, including how they are packaged with it. A new chip, it said, is a multi-year silicon effort.
The TPU 8tx is built on the same silicon as the TPU 8t1 and paired with 5th Generation AMD EPYC CPU hosts. The TPU 8i Max takes two TPU 8i chips and combines them on a single package.
They join the TPU 8t and TPU 8i, which Google introduced at Google Cloud Next in April. Google said the 8t offers up to 2.7 times better training performance and the 8i 80% better inference price performance than the previous generation of TPUs. Both are Google's own figures.
What Google says the two systems do
Google said data is normally encrypted in storage and in transit but has to be unlocked in memory while a computer processes it. The company said the TPU 8tx "builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use." The post does not describe an independent security assessment of that claim.
The 8tx also uses the Boardfly network Google created for real-time reasoning. Google said this delivers up to 2x lower latency than the TPU 8t with a "near-zero performance penalty." Google said the 8tx can fine-tune models on private data, distill large models into smaller ones and train agents through reinforcement learning on one system, without moving sensitive data between separate clusters. It described the 8tx as built for regulated enterprises and governments.
The 8i Max targets what Google calls the memory wall. Google's example is dozens of AI agents working together across an entire software codebase or hours of video, where memory bandwidth becomes the bottleneck. Google said the 8i Max delivers double the compute, double the SRAM and double the HBM bandwidth of the TPU 8i. It also said the system fits twice the silicon in the existing rack footprint, so data centers can deploy more computing power per square foot and per watt.
Benchmarks and software
Google gave no comparable performance figures for the 8i Max, and the only figure for the 8tx is the latency comparison with the 8t above. Google also said its native vLLM-TorchTPU software stack on TPUs reached "state-of-the-art price-performance marks" on the InferenceX and AgentX benchmarks. That claim covers the software stack, not the two new systems, and the post presents no independent testing.
The post also lists software changes. Google said it open-sourced TorchTPU last month1, so PyTorch code and Hugging Face models run natively on TPUs, with out-of-the-box backends for the vLLM and SGLang inference engines. It is also open-sourcing TPU Sync, a model-state layer that Google said cuts KV transfer latency and time to first token by more than 90%. Cloud Pathways, which Google described as the runtime behind Gemini, is now generally available on Google Kubernetes Engine.
The post covers further items, including a TPU agent suite in preview, an open-source XLA Serving Engine and generally available GKE dynamic slicing. This article does not detail them.
Google said TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers and universities, and that customers can now deploy TPUs directly in their own data centers. The text of Google's post available for this report does not state pricing or availability dates for the 8tx or 8i Max; the retrieved copy ends mid-sentence in its description of GKE dynamic slicing.