Google announced Trillium, its sixth-generation Tensor Processing Unit, at Google I/O on May 14, 2024. Trillium delivers a 4.7 times increase in peak compute performance per chip compared with the cost-optimized TPU v5e, while doubling both high-bandwidth memory capacity and bandwidth and doubling the inter-chip interconnect bandwidth. Google also reported that Trillium is more than 67 percent more energy-efficient than v5e, a notable claim as the power draw of AI datacenters became a growing concern.
The chip includes a third-generation SparseCore for accelerating the very large embeddings used in ranking and recommendation systems. Trillium scales to 256 chips in a single pod connected by a custom optical interconnect, and using multislice technology and Titanium offload processors it can scale further to clusters of tens of thousands of chips linked by Google’s Jupiter datacenter network.
Trillium was positioned as the hardware foundation for training and serving Google’s next generation of models. For a business reader, the headline is efficiency: as the cost of frontier AI is increasingly dominated by electricity, a generational chip whose main selling points include a two-thirds gain in energy efficiency signals where the competitive pressure in AI hardware is moving.