Google Cloud has introduced two fresh artificial intelligence (AI) chips as part of its newest tensor processing units (TPUs), marking a deeper dive into personalized hardware while still collaborating with Nvidia. The tech giant announced on April 23 that the new chips will be categorized into TPU 8t for AI model training and TPU 8i for inference, the phase where trained models handle user queries and produce results. This strategy aligns with the prevailing industry trend of optimizing hardware for specific AI tasks.
In a recent blog post, Google stated that the latest chips offer significant performance enhancements compared to prior iterations, boasting up to three times faster model training and enhanced cost-effectiveness. The company emphasized the capability to interconnect over one million TPUs into a single cluster, potentially enabling extensive computing operations with reduced energy consumption and operational expenses.
Despite this launch, Google remains committed to Nvidia’s hardware. Similar to other leading cloud service providers like Microsoft and Amazon, Google positions its personalized chips as a supplement rather than a replacement. It affirmed that it will continue to provide Nvidia’s latest processors, including the forthcoming Vera Rubin architecture, within its cloud ecosystem.
The collaboration between Google and Nvidia goes beyond hardware usage. Google mentioned ongoing joint efforts with Nvidia to enhance networking performance in data centers, including advancements to Falcon, a software-driven networking technology that Google initiated and shared through the Open Compute Project.
While Google Cloud is rolling out new AI chips to decrease its dependence on Nvidia, the chipmaker maintains market dominance with a valuation nearing $5 trillion.