Google has introduced two new versions of its in-house AI chips, known as tensor processing units (TPUs).
The TPU 8tx is designed for running AI workloads on sensitive data, while the TPU 8i Max targets large-scale applications that need more computing power and memory.
The launch brings Google’s eighth-generation TPU lineup to four systems. The company unveiled the first two, the TPU 8t and TPU 8i, at its Cloud Next event in April. Google said this is the first time it has introduced four TPU systems in a single year.
The TPU 8tx is aimed at governments and regulated businesses that need to run AI models and agents on private data.
It uses the same silicon as Google’s TPU 8t training chip and works alongside AMD’s fifth-generation EPYC processors.
Its security features are built into the hardware, keeping data and model weights isolated and protected while the chip is running.
The system can also handle smaller training jobs, allowing teams to fine-tune models on private data and run other workloads without moving that data between clusters.
Google describes the chip as offering “hardware-level confidential compute for private fine-tuning and secure agent execution”. Its faster network design also reduces latency by up to 50% compared with the TPU 8t, the company said.
The TPU 8i Max takes a different approach, combining two TPU 8i chips in a single package. Compared with the standard 8i, it doubles computing power, on-chip memory and memory bandwidth.
Google is targeting workloads where access to memory can become a bottleneck, including AI agents working across entire codebases and systems analysing hours of video.
The design also fits twice as much silicon into the same rack space, increasing the computing capacity data centres can run per square metre and per watt.
Google has not disclosed pricing or exact release dates for either chip.
Amin Vahdat, the company’s chief technologist for AI and infrastructure, wrote in the blog post that Google’s TPU business now has nearly 100 customers. They include frontier AI labs, trading firms, retailers and universities.
TPUs are among the main alternatives to Nvidia’s chips for AI workloads. Google makes them available through its cloud business and works with four design partners to develop them.
Published · Updated
Back to top