Google Announces 6th Generation Trillium TPU: Performance, Memory Bandwidth, Efficiency Improvements

Published by

During the I/O 2024 Developers Conference, Google unveiled its 6th generation data center AI chip, the Tensor Processing Unit (TPU) known as Trillium. Google CEO Sundar Pichai announced that the Trillium TPU would be available later this year. The latest TPU iteration promises a 4.7-fold increase in computational performance compared to its predecessor. This improvement is attributed to expanded matrix multiplication units (MXU) and increased overall clock speeds. Additionally, the memory bandwidth of the Trillium chip has doubled, enhancing its ability to handle larger data sets and complex computations more efficiently.

Trillium incorporates the third generation of Google's SparseCore technology, which Google describes as a specialized accelerator designed to manage common large-scale tasks in advanced ranking and recommendation workloads. This enhancement is expected to expedite the training of models and reduce latency in model deployment.

Pichai emphasized that the new Trillium TPU is Google's most energy-efficient to date, a critical advancement given the exponential growth in demand for AI chips. Over the past six years, the industry’s need for machine learning computation has increased by a factor of one million, roughly doubling each year. Without reducing the power consumption of these chips, sustainability would be challenging. To address this, Google has achieved a 67% reduction in energy consumption with the Trillium TPU compared to the fifth generation. This launch is part of Google's ongoing efforts to lead in the GPU and AI processor markets, continuously pushing the boundaries of what is possible in AI hardware advancements.


Share this content
Twitter Facebook Reddit WhatsApp Email Print