Google Cloud Unveils Trillium, Its 6th-Gen TPU With A Huge 4.7X AI Performance Leap

Google took the wraps off its latest Tensor Processing Unit (TPU), Trillium, at this year’s Google I/O conference. Trillium is custom silicon designed by Google do handle AI-specific tasks. According to the company, these sixth-generation accelerators “achieve an impressive 4.7X increase in peak compute performance per chip compared to TPU v5e.” Trillium will play an important role in powering Google AI models such as Gemini 1.5 Flash, Gemma 2, and Imagen 3.

Google says it was able to obtain this leap in performance because Trillium uses twice the amount of High Bandwidth Memory (HBM) in terms of both bandwidth and capacity. Moreover, Trillium makes use of the company's 3rd generation SparseCore technology, which Google describes as a specialized accelerator for processing recommendations and advanced ranking workloads. These developments lead to energy efficiency gains of over 67% compared to the prior generation.


Trillium is one piece of the important puzzle that comes together to make what Google has dubbed its AI Hypercomputer. These TPUs work in concert with GPUs, storage systems, custom networking and open-source frameworks to deliver the performance necessary for complex AI workloads running in the cloud. Google says that these new chips will also be able continue to work with models built to run on prior generation accelerators, which is a big win for developers, in terms of backwards compatibility.

It's good to see Google be able to achieve these kinds of efficiency gains. It helps address two of the issues currently facing AI. The first is keeping the cost of scaling and running AI workloads under control as explosive demand for the technology continues to accelerate. Second, this should go along way toward minimizing the amount of power required to run these workloads in support of greener solutions.

As exciting as these upgrades might be for developers, patience will be needed. Google didn’t give an exact date as to when its new Trillium TPU technology and instances powered by it will be available, only saying that it will become available “later this year.”