Google Cloud Launches TPU 8t and 8i AI Chips
Google Cloud revealed on Wednesday its eighth generation of custom AI chips, known as tensor processing units or TPUs. The company divided this generation into two versions. The TPU 8t targets model training. The TPU 8i focuses on inference, which involves the continued operation of models after users provide prompts.
Performance Gains Over Prior Generations
Google highlights strong specs for these TPUs when measured against earlier versions. Training for AI models runs up to three times faster. Customers receive 80 percent better performance for each dollar spent. Engineers can link more than one million TPUs into one cluster. These advances deliver greater computing power while using less energy and lowering costs for users. Google chose the TPU name for these custom, energy-efficient chips, a label that dates back to their origins as Tensor processors.
Google Cloud, a major player in cloud services since its founding in 2008, has long invested in specialized hardware for AI tasks. TPUs first appeared publicly in 2016, designed to handle machine learning workloads more efficiently than general-purpose processors.
Balancing Competition and Partnership with Nvidia
These new chips do not represent a direct challenge to Nvidia's position. Google, along with other large cloud operators like Microsoft and Amazon, uses its TPUs to add to the Nvidia-powered setups already available. The company has no plans to remove Nvidia hardware. Google confirmed it will offer Nvidia's next chip, Vera Rubin, through its cloud later this year.
Cloud giants such as Amazon, Microsoft, and Google continue to develop their own AI processors. Over time, businesses may shift more AI operations to these clouds and adapt software for the custom chips. This could reduce reliance on Nvidia as customers consolidate workloads.
Nvidia's Enduring Strength
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Experts caution against underestimating Nvidia. Chip analyst Patrick Moore shared a lighthearted note on X. In 2016, he suggested Google's initial TPU launch might hurt Nvidia and Intel. Nvidia has since grown to a market value approaching five trillion dollars, proving that forecast wrong over the years.
Nvidia expects benefits from providers like Google expanding AI services. More demand for cloud AI could mean additional orders for Nvidia products, even if some tasks run on Google hardware.
Google Cloud also committed to joint efforts with Nvidia. The pair aims to improve networking for Nvidia systems in Google's environment. They focus on enhancing Falcon, a software networking solution Google developed and released as open source in 2023. This work falls under the Open Compute Project, a key group for data center hardware standards.
Nvidia dominates the AI accelerator market with its GPUs, powering most training and inference today. Partnerships like this one help cloud providers scale efficiently while maintaining compatibility.
Broader Context in AI Hardware Race
The announcement came during Google Cloud Next, an event showcasing updates in cloud and AI. As AI demand surges, providers balance in-house innovation with industry-standard components. Google's approach supports diverse workloads, from training massive models to running inferences at scale.
This strategy positions Google Cloud to attract enterprises building AI applications, offering choices in hardware without forcing migrations.