Google officially announced that its sixth-generation Tensor Processing Unit (TPU), Trillium, is now available for rent. After months of being offered in preview, this powerful AI chip is ready for general use. Designed to revolutionise AI infrastructure, Trillium has already proven its capabilities by training Gemini 2.0, Google’s cutting-edge AI model.
This news is big if you’re in the AI or tech industry. The chip boasts impressive features, including double the High Bandwidth Memory (HBM) capacity and double the Interchip Interconnect bandwidth compared to its predecessors. According to Google, Trillium offers up to a 2.5x improvement in training performance per dollar, making it an efficient choice for businesses aiming to optimise their AI operations.
Revolutionary performance with Trillium TPU
Trillium isn’t just an upgrade—it’s a leap forward. It delivers more than four times the training performance of its predecessor, while energy efficiency has increased by 67%. Regarding raw computing power, each chip’s peak performance is 4.7 times greater than earlier.
Google’s benchmarks reveal that Trillium also significantly enhances inference tasks. For image generation models like Stable Diffusion XL, throughput has increased thrice. Large language models, crucial in today’s AI landscape, see nearly double the throughput.
The chip’s architecture is also optimised for embedding-intensive models. Its third-generation SparseCore improves dynamic and data-dependent operations efficiency, ensuring smooth performance even under complex workloads.
Powering Google Cloud’s AI Hypercomputer
One of Trillium’s standout achievements is its role in Google Cloud’s AI Hypercomputer. This advanced system integrates over 100,000 Trillium chips, all connected through a Jupiter network fabric with an astounding 13 Petabits/sec bandwidth. The system combines this cutting-edge hardware with open-source software and well-known machine-learning frameworks like JAX, PyTorch, and TensorFlow.
What does this mean for you? Google Cloud customers can now use the same state-of-the-art hardware that trained the Gemini 2.0 AI model. With Trillium’s general availability, high-performance AI technology is no longer reserved for a select few. From image generation to complex language models, Trillium opens the door to countless applications, making it a valuable asset for businesses aiming to stay ahead in the AI race.
