Google has unveiled the sixth generation of its custom Tensor Processing Unit (TPU) AI chip, dubbed Trillium.
Announced at the company’s annual I/O developer conference in California, Trillium has been designed to support large language and recommender models, including Google’s Gemini, Imagen, and Gemma, with reduced latency and lower costs.
Described by the company as its “most advanced TPU," Google said this latest iteration can achieve a 4.7× increase in peak compute performance per chip compared to TPU v5e. The company said this has been achieved by expanding the size of the chip’s matrix multiple units (MXUs) and increasing the clock speed.
It also comes equipped with a third-generation SparseCore – a specialized dataflow accelerator that processes large embeddings often found in ranking and recommender systems.
Trillium contains double the High Bandwidth Memory (HBM) capacity and bandwidth of its predecessor and doubles the Interchip Interconnect (ICI) bandwidth. Google claims it is 67 percent more energy-efficient than TPU v5e.
The processor can also scale up to 256 TPUs in a single pod, which can then scale to hundreds of pods using the company’s multislice technology introduced last year. Google said this allows for “tens of thousands” of chips to be connected via Google's Jupiter data center network technology.
Google also announced that Trillium TPUs will form part of Google Cloud's AI Hypercomputer, which debuted in December 2023. The cloud-based supercomputer architecture was designed for AI workloads and combines performance-optimized hardware, open software, machine learning frameworks, and flexible consumption models.
“Google Cloud TPUs are the cutting-edge of AI acceleration, custom-designed and optimized to empower large-scale artificial intelligence models,” Amin Vahdat, VP and GM of ML, Systems, and Cloud AI at Google, said in a blog post announcing Trillium.
“Trillium TPUs will power the next wave of AI models and agents, and we’re looking forward to helping enable our customers with these advanced capabilities,” he added.