Google Introduces Trillium, the Latest Generation of TPUs Designed for AI

At its Google I/O conference, the tech giant unveiled a number of major language models such as Gemini 1.5 Flash, Imagen 3 and Gemma 2. What they have in common: all of these models were trained on TPUs ( Tensor Processing Unit). Believing that the pace of innovation is accelerating, Google is taking the lead and this week announced Trillium, sixth generation TPUs, described as the highest performance and most energy efficient to date.

Some figures announced by Google offer an overview of the capabilities of these specialized chips to accelerate calculations linked to artificial intelligence. Trillium TPUs show a 4.7x increase in peak computing performance per chip compared to fifth-generation TPUs. The firm says it has doubled the capacity and bandwidth of high-bandwidth memory (HBM), as well as the bandwidth of the inter-chip interconnect (ICI) compared to its predecessors.


Energy savings highlighted

Google also indicates that SparseCore processors accelerate heavy workloads to integrate by strategically offloading random access from TensorCore, explains the firm. And the expected result at the turning point is the following: “Trillium TPUs make it possible to train the next wave of baseline models faster and serve those models with reduced latency and lower cost,” adds Google.

Another advantage: Trilliums are 67% more energy efficient than fifth generation TPUs. Obviously, we will have to wait for the official launch to confirm or not the veracity of the figures put forward by Google.

Several confirmed beta testers

In the meantime, the giant is certain: “Trillium TPUs will power the next wave of AI models and agents.” Several companies have already indicated that they are ready to exploit the potential of Trillium, notably the company Essential AI, specializing in autonomous vehicles, but also Nuro, a robotics specialist which plans to train its models with Cloud TPUs, or again Deep Genomics which is working on drug discovery using AI.

The TPUs will also be used internally since Google DeepMind will use them to train and deploy the next generations of LLM Gemini “faster, more efficient and with lower latency”, promises Google.


Hugging Face, partner in the provision of these TPUs

“Our partnership with Google Cloud makes it easier for Hugging Face users to refine and run open models on Google Cloud's AI infrastructure, including TPUs. We will make Trillium's performance available to all AI developers thanks to our new Optimum-TPU library”, specifies Jeff Boudier, product manager at Hugging Face.

Do you want to stay up to date on the latest news in the artificial intelligence sector? Register for free to the IA Insider newsletter.

Selected for you