Alphabet's Google recently demonstrated that their tensor processing unit is a force to be reckoned with in computing. They benchmarked their TPUs and posted an astounding rate of 15 times to 30 times faster processing times. This is compared to an Intel Haswell central processing unit (CPU)/Nvidia K80 graphics processing unit (GPU) combination.
Google first announced that they were working on chips in-house back in May 2016 during their I/O developer conference. They revealed that these chips will be called tensor processing units and will be built to complement their TensorFlow machine-learning framework. Aside from those points, they did not offer any more information, TechCrunch noted.
Last week, the Alphabet subsidiary offered a more intimate and detailed look at their tensor processing units. Its processing speed is most worthy of note that was benchmarked against Intel CPU and Nvidia GPU. The litmus test showed that Google's TPUs are capable of processing machine learning workloads 15 times to 30 times faster. Furthermore, the company's in-house chips can also provide 30x to 80x higher performance/watt that could potentially go higher with faster memory in the future.
Those numbers imply that Intel and Nvidia have a potential rival at their doorsteps. The former has been monopolizing the chip industry for decades, but that may change soon. Cloud computing needs are changing and it looks like Google's tensor processing unit is better poised to address those needs.
Seeking Alpha pointed out that Google's TPU benchmark is not cutting edge because it was conducted in 2015. Since then, both Intel and Nvidia have come up with better-performing chips. Nevertheless, the tech giant wrote in their paper that the tensor processing unit could be further improved by them or by others.
Google first thought about tensor processing units back in 2006. However, there was no need at the time for such a technology. But in 2013, the Alphabet subsidiary embarked on a project with this objective: to create a processor that could outrun GPUs 10x. It took them three years to announce the project to the public, though, and even then they did not offer a lot of details.
