Google says its custom-built tensor processing units (TPUs) run the core numerical workloads behind virtually all of its products and services, from Search and YouTube recommendations to the Gemini family of AI models.
The company designed the chips "from the ground up more than a decade ago" to handle the large-scale matrix and tensor calculations that AI models require, an approach that set it apart from rivals relying on off-the-shelf graphics processing units (GPUs) from Nvidia and AMD.
Where GPUs were originally designed for rendering graphics and later adapted for AI, TPUs were purpose-built from the outset for machine learning, optimised specifically for the kind of repetitive, high-volume arithmetic that underpins model training and inference.
Google says the specialisation allows its chips to shorten model runtimes by performing those calculations more efficiently than general-purpose hardware.
The latest generation represents a significant step up in raw performance.
"The newest generation of TPUs can process 121 exaflops of compute power with double the bandwidth of previous generations," the company said, a reference to the TPU 8t training chip unveiled at the Cloud Next conference in Las Vegas earlier this week.
Google also describes its processors as compact and energy-efficient, a design priority as the power consumption of AI data centres has become a growing concern across the industry.
The company's decision to develop its own silicon rather than depend entirely on third-party suppliers has given it greater control over its AI infrastructure stack, from chip design through to the software frameworks and cloud services that sit on top.
Related reading
- Google in talks with Marvell to develop two new AI chips as it builds multi-supplier TPU strategy
- Nvidia launches open-source quantum AI models as it deepens grip on chip design workflows
- Meta partners with Broadcom on MTIA AI chips
That vertical integration now extends across two distinct chip lines, with the TPU 8t optimised for training and the TPU 8i built for inference and agentic workloads, each tailored to different stages of the AI pipeline.
Google's TPU programme has also become a commercial offering through Google Cloud, where outside companies can rent access to the chips, and major AI laboratories including Anthropic have signed large-scale capacity agreements for next-generation TPU access.
The recap
- Google highlights its custom Tensor Processing Units powering products.
- Newest TPU generation processes 121 exaflops compute power.
- Company directs readers to an explanatory video for more.
