Google announces Trillium TPU for its Cloud, which is the most advanced and energy-efficient hardware for powering the next generation AI models. Trillium is already being used by Google to run its most advanced AI model Gemini 1.5 Pro. Trillium is a custom AI-specific hardware from Google that is being used by Google for running its latest generative AI models like Gemini 1.5 Flash, Imagen 3 and Gemma 2.0. These models are trained on the Trillium TPU and the same TPU is being used to serve these models.

During its I/O 2024 event on Tuesday, Google announced the 6th generation Google TPU (Tensor Processing Units) called Trillium. According to the company statement, Gemini 1.5 Pro is the company's largest and most capable AI Model trained on the Trillium GPUs.

Google used tens of thousands of Trillium GPUs for training its most advanced Gemini 1.5 Pro models. Company used Trillium GPUs for training Gemini 1.5 Flash, Imagen 3, and Gemma 2.0 models.

Trillium GPU is 4.7X faster in peak compute per as compared to previous generation v5e TPU, so Trillium is much capable of running next generation Generative AI models. The Trillium GPU also comes with the high bandwidth memory (HMB) and double the Interconnect (ICI) bandwidth. The Trillium TPU is baked on the third-generation SparseCore, which is a special accelerator for processing ultra-large embeddings. So this GPU gives much better performance in advanced ranking and recommendation ML workloads.

According to Google Trillium TPU can be used to train the next wave of AI models faster and with low cost. These TPU are also over 67 per cent more energy efficient as compared to its predecessor TPUs. The Trillium TPUs will bring the next wave of various AI models trained in an energy efficient way.

According to Google, Trillium GPU can train the next generation AI models faster and also with reduced latency. The Trillium GPU is more energy efficient compared to its predecessor, so the training cost will be much more energy efficient.

Trillium can be used to make a high computing environment for training and serving models. Trillium can be scaled up to 256 TPUs in a single low-latency pod with a high-bandwidth connectivity. This technology enables Google to interconnect thousands of chips to build a supercomputer capable of processing petabytes of data in a second.

Trillium is Google’s sixth generation of Tensor TPU which is 4.7 times faster than the previous TPU. CEO Sundar Pichai said that the new Trillium chips are coming later this year. Being 4.7 times faster than their predecessors, Trillium is going to train and infer the next generation of AI models.

Google also announced it will provide Nvidia’s Blackwell GPU starting in 2025 on its cloud.  Company further said that they will be “one of the first” cloud companies to offer Nvidia’s Blackwell GPU.

Google first introduced its first TPU v1 bank in the year 2013, followed by the next release of its TPU in 2017. Now the company is releasing Trillium TPUs which are much faster and capable of training models fast and with energy efficiencies.

As of now Trillium is a part of Google’s AI Hypercomputer, a supercomputer architecture, which is designed for running cutting-edge AI workloads. Google is partnering with Hugging Face  for the optimization of hardware for open-source model training and serving.

Features of Trillium GPU

? 4.7 times faster  than their predecessors

? Low energy consumptions

? Well designed for next generation model training

? Powerful GPU for model serving

? Trillium can be scaled up to 256 TPUs

? 67 per cent more energy efficient