Google Cloud has formally introduced the final availability (GA) of its sixth-generation Tensor Processing Unit (TPU), often known as Trillium. Based on the corporate, the AI accelerator is designed to satisfy the rising calls for of large-scale synthetic intelligence workloads, providing extra efficiency, vitality effectivity, and scalability.
Trillium was introduced in Could and is a key part of Google Cloud’s AI Hypercomputer, a supercomputer structure that makes use of a cohesive system of performance-optimized {hardware}, open-source software program, main machine studying frameworks, and adaptable consumption fashions.
With the GA of Trillium TPUs, Google enhanced the AI Hypercomputer’s software program layer, optimizing the XLA compiler and common frameworks like JAX, PyTorch, and TensorFlow for higher value efficiency in AI coaching and serving. Options like host-offloading with massive host DRAM complement Excessive Bandwidth Reminiscence (HBM) for improved effectivity.
The corporate states that Trillium delivers coaching efficiency over 4 occasions and as much as thrice the inference throughput in comparison with the earlier era. With a 67% enchancment in vitality effectivity, Trillium is quicker and greener, aligning with the growing emphasis on sustainable expertise. Its peak compute efficiency per chip is 4.7 occasions increased than its predecessor, making it appropriate for computationally intensive duties.
Trillium TPUs had been additionally used to coach Google’s Gemini 2.0 AI mannequin, with a correspondent on a Hacker Information thread commenting:
Google silicon TPUs have been used for coaching for at the very least 5 years, most likely extra (I believe it is 10 years). They don’t rely upon Nvidia GPUs for almost all of their initiatives. It took TPUs some time to make amends for some particulars, like sparsity.
That is adopted by a remark that notes that TPUs have been used for coaching deep prediction fashions in advertisements since at the very least 2018, with TPU capability now probably surpassing the mixed capability of CPUs and GPUs.
Presently, Nvidia holds between 70% and 95 % of the AI knowledge middle chip market, whereas the remaining proportion contains totally different variations like Google’s TPUs. Google doesn’t promote the chips straight however presents entry via its cloud computing platform.
In a Reddit thread, a correspondent commented concerning not promoting the chips:
That is proper, however I believe Google is extra future-focused, and environment friendly AI will in the end be rather more useful than chips.
In my nation, we frequently say that we must always make wooden merchandise relatively than export wooden as a result of making furnishings creates extra worth. I believe that is comparable: TPUs and AI create extra worth than the 2 issues alone.
Extra particulars on pricing and availability can be found on the pricing web page.
