Google’s parent company, Alphabet, has launched its most advanced AI chip, Trillium. This sixth-generation Tensor Processing Unit (TPU) promises a 67% boost in energy efficiency compared to its predecessor, TPU v5e. Trillium is designed to supercharge Google’s cloud computing platform with unparalleled performance and efficiency.
Trillium’s standout features include a 4.7X increase in peak compute performance, doubled High-Bandwidth Memory (HBM) capacity, and enhanced Interchip Interconnect (ICI) bandwidth. The chip also incorporates a third-generation SparseCore, optimising it for ultra-large embeddings common in ranking and recommendation workloads.
A significant edge for Trillium is its scalability. It can integrate up to 256 TPUs within a single high-bandwidth, low-latency pod, and scale to hundreds of pods with multislice technology and Titanium Intelligence Processing Units (IPUs). This enables a supercomputer configuration interlinked by a multi-petabit-per-second datacenter network.
Google’s blog highlights that Trillium TPUs will revolutionise AI models, powering the next wave of AI agents with enhanced capabilities. This new chip will aid companies like Deep Genomics and Deloitte, as well as Google DeepMind’s Gemini models, by offering faster, more efficient, and lower latency training and serving.
The Trillium TPUs are a cornerstone of Google Cloud’s AI Hypercomputer, designed for cutting-edge AI workloads. With Trillium, Google continues to challenge Nvidia’s dominance in the AI chip market, offering a robust alternative for high-performance AI processing.
