Edge TPUs revealed as the future of AI for Google at Next 2018
Google had previously emphasized its commitment to tensor processing units (TPUs) for its 'brand' of AI. Accordingly, attendees at the Next 2018 conference, an event mainly for enterprise-focused Google Cloud, were introduced to a new form of this technology: the Edge TPU.
The TPU is Google's take on the application-specific integrated circuit (ASIC). The ASIC, in turn, is a parallel-processing-oriented variation on a GPU for the purposes of neural network training. ASICs differ from GPUs in that they also supply libraries and instructions that allow the processor to operate on local data. Therefore, the ASIC may accelerate the training process compared to the GPU.
Google's existing TPUs run the company's proprietary TensorFlow training software in order to set up the desired neural network. The Edge TPU, on the other hand, is simply intended to carry out the tasks in question. This will enable Google to provision more AI components - i.e. software, data, and now both operating and executing hardware. In other words, it makes Google's AI services more end-to-end. An early example of Edge TPUs in action is found on LG production lines, where the units check TV panels for flaws.
Google also debuted its on-premises Google Kubernetes Engine (GKE On-Prem), which is intended to unify the data in a company's center and cloud environments. It is pretty clear from the Next 2018 sessions that Google wants to take over the world of enterprise data.