At I/O Google took the wraps off its sixth Tensor Processing Unit (TPU), codenamed Trillium, designed to support a new generation of larger, more capable large-scale language and recommendation models.

Originally built to accelerate Google's internal machine learning workloads, such as those built into Gmail, Google Maps and YouTube, the search giant began making the matrix math accelerators available in its cloud in 2018.

Six generations later, Google's TPUs are central to the development of Gemini's large language models behind its growing portfolio of generative AI apps and services.

According to Google, Trillium has 4.7 times the compute performance and double the memory capacity and high-bandwidth bandwidth compared to its previous TPU v5e design, which we looked at last summer. Google has also doubled the interchip bandwidth.