Google Cloud has announced the general availability of Trillium, its latest Tensor Processing Unit (TPU) designed for large-scale AI and machine learning workloads. Trillium TPUs offer significant performance improvements over previous generations, with up to 1.5 petaflops of AI compute power per chip. They are optimized for advanced AI models like large language models, generative AI, and multimodal AI. Key features include a high-performance unified cache architecture, enhanced matrix multiplication capabilities, and support for bfloat16 data types. Trillium TPUs are available on Google Cloud’s AI-optimized VMs and can be deployed in private, public, or hybrid cloud environments. They aim to provide researchers and developers with the computational power needed to push the boundaries of AI innovation. The announcement highlights Google’s continued investment in specialized hardware for AI acceleration.
Source: https://cloud.google.com/blog/products/compute/trillium-tpu-is-ga