Google expands 8th Gen Tensor series with agents workload-focused AI chips

Google adds TPU 8tx and TPU 8i Max, targetting secure AI inference and memory-intensive workloads, as company expands its custom chip portfolio to support enterprises deploying autonomous AI agents

Google has expanded its eighth-generation Tensor Processing Unit (TPU) portfolio with two new AI chips: TPU 8tx and TPU 8i Max. According to the US-based technology giant, these chips are designed to support the growing demands of agentic artificial intelligence (AI). The new systems are dedicated to enterprises that run AI agents handling sensitive data, process larger volumes of information and perform complex, multi-step reasoning tasks.

 

The launch also expands Google's efforts to make its in-house AI chips accessible to developers and businesses as an alternative to conventional graphics processing units (GPUs, which comes from the house of Nvidia.

 

According to Google, the two systems address different computing requirements. While TPU 8tx focuses on secure and confidential AI inference, TPU 8i Max is designed to handle large context windows and memory-intensive workloads.

 

TPU 8tx: Focus on confidential AI computing

Google said that the TPU 8tx is designed for enterprises and government organisations that need to run AI models and autonomous agents alongside sensitive information. It focuses on confidential computing, allowing data and model weights to remain protected even while they are being processed.

 

Typically, data is encrypted when stored or transferred across networks but needs to be decrypted in memory when a computer processes it. Google said the TPU 8tx addresses this gap by building security into the hardware layer.

 

The system uses the same silicon as TPU 8t and is paired with fifth-generation AMD EPYC CPU hosts. It also uses Google's Boardfly network, which is designed to reduce communication delays during real-time AI reasoning.

 

According to Google, TPU 8tx can deliver up to two times lower latency compared with TPU 8t, with almost no performance penalty.

 

The system also allows organisations to run and customise AI models on the same infrastructure. This includes fine-tuning models using private datasets, distilling larger models into smaller ones and training AI agents through reinforcement learning without moving sensitive data between separate computing clusters. 

TPU 8i Max: Designed for larger AI workloads

The second system, TPU 8i Max, is designed to address memory limitations when multiple AI agents work together or process large volumes of information.

 

Google said memory bandwidth can become a bottleneck when AI agents work across entire software codebases or analyse hours of video simultaneously.

 

To address this, the company has combined two low-latency TPU 8i chips into a single package. The resulting system offers twice the computing power, twice the static random-access memory (SRAM) and twice the high-bandwidth memory (HBM) bandwidth of TPU 8i.

 

The company said the configuration allows processors to access data more quickly and reduces the chances of large workloads slowing down because of memory constraints. It also fits twice the silicon within the existing rack footprint, allowing data centres to deploy more computing capacity per square foot and per watt.

 

Together, the four systems in Google's eighth-generation portfolio target different stages of AI development. TPU 8t is designed for large-scale model training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing and TPU 8i Max for workloads requiring larger context windows.

 

Alongside the hardware announcement, Google is introducing software tools to make TPUs easier for developers to use, including support for PyTorch, Hugging Face models and open-source inference engines. The updates also aim to reduce AI infrastructure costs, improve memory management and make large-scale TPU workloads easier to deploy.