[DiaryPoint | Reporter Eom Ju-yong] Google announced its eighth-generation Tensor Processing Unit (TPU) systems, 8tx and 8i Max, on November 9. The 8tx is designed for confidential inference, protecting sensitive data while it is being processed. The 8i Max is built for AI workloads that need to handle large contexts at once.
Rather than entirely new chip designs, both systems build on the existing 8t and 8i chips, with changes to memory, networking and security configurations. Google introduced the 8t in April for training large AI models and the 8i for real-time inference. The latest announcement expands the eighth-generation lineup to four systems.
8tx isolates data in use at the hardware level
The 8tx uses the same silicon as the 8t and pairs it with a fifth-generation AMD EPYC central processing unit (CPU) as its host. Data is encrypted while stored or in transit, but must be decrypted in memory to be processed. The 8tx uses confidential computing to protect that data in use, along with model weights, at the hardware level.
AMD said it extends the CPU’s security feature, Secure Encrypted Virtualization (SEV), to the TPU accelerator. Before allowing access to protected memory, the system verifies the accelerator’s identity and security configuration. It encrypts data in CPU memory and as it moves between the CPU and accelerator, then provides hardware isolation while the accelerator processes it. Google said the 8tx’s Boardfly network can cut latency by up to two times compared with the 8t, with almost no performance loss.
8i Max combines two chips to handle longer contexts
The 8i Max combines two low-latency TPU 8i chips in a single package. According to Google, it doubles compute, SRAM capacity and HBM bandwidth compared with the 8i, while fitting twice as much silicon into the same rack footprint to increase compute density per unit of space and power.
When multiple AI agents analyze long inputs together—such as an entire software codebase or hours of video—memory bandwidth can become a bottleneck as they fetch the data they need. The 8i Max is designed to ease that constraint in tasks involving large contexts, including coding assistance and complex reasoning.
Google uses TPUs in its own services, including Search, YouTube and Gemini, and also offers them to Google Cloud customers. Reuters reported that Google Cloud revenue more than doubled to $30.71 billion in the September quarter, with TPU sales making a meaningful contribution to results. Google is positioning its systems for sensitive AI workloads in regulated industries and government, as well as businesses working with large contexts, and is expanding the options available in a market dominated by Nvidia.