Google alerts Nvidia by unveiling its Trillium artificial intelligence chip and promises to launch the H200 Tensor Core GPUs in a few days.

Google Cloud has introduced Trillium, its new sixth-generation TPU processor designed to handle advanced artificial intelligence workloads. Initially announced in May 2024, Trillium focuses on delivering improved performance and greater cost efficiency for large-scale training, fine-tuning, and inference tasks.

This innovation is part of Google Cloud's AI Hypercomputer infrastructure, which combines TPUs, GPUs, and CPUs with open-source software to meet the growing demand for generative artificial intelligence. Trillium brings significant improvements over its predecessor, the TPU v5e, achieving more than a 4x increase in training performance and up to a 3x increase in inference capacity. Additionally, it offers double the HBM capacity and a significant increase in Interchip Interconnect (ICI) bandwidth, making it ideal for large language models such as Gemma 2 and Llama, as well as computationally intensive inference applications, including diffusion models like Stable Diffusion XL.

Another key aspect of Trillium is its energy efficiency, with a 67% improvement compared to previous versions, a factor also highlighted by Google. In performance tests, Trillium has demonstrated training speeds up to 4 times faster for models such as Gemma 2-27b and Llama2-70B, while in inference tasks, it has shown performance 3 times greater than that of the TPU v5e, especially in models that require a high level of computation.

Scalability is another significant advantage of Trillium. This TPU can connect up to 256 chips in a single high-capacity pod, with the potential to expand to thousands of chips within Google’s Jupiter data center network, providing nearly linear scalability for extensive training tasks. Thanks to the Multislice software, Trillium ensures consistent performance across hundreds of pods.

In addition to Trillium, Google has launched the A3 Ultra VMs that feature Nvidia H200 Tensor Core GPUs. These high-performance virtual machines will be available in preview this month, offering Google Cloud customers a robust option within the company’s artificial intelligence infrastructure.