With the unveiling of the new Trillium TPUs at the Google I/O event, artificial intelligence models are set to operate faster and more efficiently. Google is taking significant steps in the hardware realm to meet the high processing power, memory, and communication capacity required by the rapidly evolving productive AI technologies. At the Google I/O 2024 event, the company announced Trillium, the most powerful and energy-efficient Tensor Processing Unit (TPU) to date.
Trillium: A Groundbreaking Step in Performance and Efficiency
Trillium TPUs deliver a 4.7x increase in per-chip processing performance compared to TPU v5e. Google has doubled the High Bandwidth Memory (HBM) capacity and bandwidth and also doubled the Inter-Chip Interconnect (ICI) bandwidth compared to TPU v5e.
Equipped with a third-generation SparseCore, Trillium comes with a specialized accelerator to handle ultra-large embeddings commonly used in advanced ranking and recommendation workflows.
Trillium TPUs enable faster training of AI models and deliver these models with lower latency and cost. The 6th generation of TPUs is also the most sustainable: Trillium TPUs offer over 67% more energy efficiency compared to TPU v5e.
Scalable up to 256 TPUs in a single high-bandwidth, low-latency partition, Trillium goes beyond pod-level scalability. With multislice technology and Titanium AI Processing Units (IPUs), Trillium TPUs can scale to hundreds of pods, connecting tens of thousands of chips with a multi-petabit-per-second data center network, creating a building-scale supercomputer.
Pioneering AI-Focused Hardware
For over a decade, Google has been pushing the boundaries of scale and efficiency in AI by developing specialized hardware. Starting with the development of the world’s first purpose-built AI accelerator, TPU v1, in 2013, Google followed up with the first Cloud TPU in 2017.
Without TPUs, Google’s most popular services such as real-time voice search, photo object recognition, interactive language translation, and state-of-the-art foundational models like Gemini, Imagen, and Gemma would not be possible. The scale and efficiency of TPUs enabled fundamental research on Transformer architectures, which underpin modern productive AI.
Trillium and AI Hypercomputer
Trillium TPUs are part of AI Hypercomputer, a groundbreaking supercomputer architecture designed specifically for cutting-edge AI workloads. AI Hypercomputer integrates performance-optimized infrastructure, including Trillium TPUs, open-source software frameworks, and flexible consumption models.
Supporting open-source libraries like JAX, PyTorch/XLA, and Keras 3, Google empowers developers and facilitates model training and deployment. Google also partnered with Hugging Face to facilitate model training and deployment on Optimum-TPU.
AI Hypercomputer also offers flexible consumption models for AI and machine learning workloads. Dynamic Workload Scheduler (DWS) addresses access to AI and machine learning resources, helping customers optimize their spending.
The flexible boot mode improves the experience of sudden workloads like training, fine-tuning, or batch jobs by allowing you to program all required accelerators simultaneously, regardless of your entry point.
Powering the Next Wave of AI Innovation
Trillium TPUs will empower the next wave of AI models and tools, and Google is excited to offer these enhanced capabilities to its customers. Autonomous vehicle company Nuro is committed to creating a better daily life through robotics by training its models with Cloud TPUs.
Deep Genomics is strengthening the future of drug discovery with AI, and they are eagerly anticipating how their next foundational models supported by Trillium will change patients’ lives. Deloitte, named Google Cloud’s AI Partner of the Year, will deliver Trillium to transform businesses with productive AI support.
Trillium TPUs are the result of over a decade of research and innovation and will be available for sale later this year. Google aims to usher in a new era of AI innovation with Trillium.”
