Google have announced that Trillium, its sixth-generation Tensor Processing Unit (TPU), is now available in preview to Google Cloud customers. The new chip offers significant improvements in performance and energy efficiency compared to its predecessor.
The new Trillium processor delivers 4.7x peak compute performance per chip compared to the prior generation, TPU v5e, while achieving 67% greater energy efficiency.
Initially revealed at Google I/O in May, these specialised processors are found exclusively in Google data centres, powering AI services across Google's products, including Search, YouTube, and DeepMind's large language models.
The processors represent three different levels of specialisation:
- CPUs (Central Processing Units): General-purpose chips found in smartphones and laptops
- GPUs (Graphics Processing Units): Processors for accelerated compute tasks, found in high-end gaming systems or some desktop devices
- TPUs (Tensor Processing Units): Google's custom-designed chips specifically for AI computations, found only in Google data centres
Google started thinking about TPUs about 10 years ago when they realised that if every user started "talking" to Google for just three minutes a day, they would need to double the number of computers in their data centres.