Google Cloud has introduced two novel artificial intelligence (AI) chips as part of its latest tensor processing units (TPUs) generation, showcasing a deeper commitment to tailor-made hardware while still collaborating with Nvidia.
Announced on April 23, the new chips come in two variants: the TPU 8t, optimized for AI model training, and the TPU 8i, tailored for inference tasks where trained models handle user queries and generate results. This strategy aligns with the industry trend of fine-tuning hardware for specific AI functions.
In a recent blog post, Google emphasized that the fresh chips deliver substantial performance enhancements compared to previous iterations, boasting up to three times faster model training speeds and enhanced cost-effectiveness. Furthermore, Google highlighted the capability to interconnect over one million TPUs in a single cluster, potentially facilitating large-scale computing operations with reduced energy consumption and operational expenses.
Despite the chip launch, Google remains committed to Nvidia hardware. Similar to other major cloud service providers like Microsoft and Amazon, Google positions its custom chips as a supplement rather than a replacement. The company confirmed its continued support for Nvidia’s latest processors, including the upcoming Vera Rubin architecture, within its cloud infrastructure.
The collaboration between Google and Nvidia extends beyond hardware. Google disclosed ongoing joint efforts with Nvidia to enhance data center networking performance, with a focus on refining Falcon, a software-driven networking technology developed and shared via the Open Compute Project.
While Google Cloud’s introduction of new AI chips aims to reduce dependency on Nvidia, the chipmaker maintains its market dominance, with an estimated valuation nearing $5 trillion.
