Artificial Intelligence in Buyer’s Hands


Artificial Intelligence in Buyer Hands has unveiled a star Pixel smartphone. That is powered by its first mobile chip. This is an amazing step to put artificial intelligence in buyer’s hands. Google is driving the focus on the new system on a chip (SoC). That would be inside the new Pixels. This is named as the Tensor SoC. It is called after the Tensor Processing Units (TPU) Google uses in its data centers to allow computers to think more like people do.

It’s mainly a mobile system on a chip planned around artificial intelligence. Google’s Pixel line has taken a limited share in a worldwide smart phone market. That is dominated by Samsung, Apple and Chinese manufacturers. Pixel phones have been understood as a method for Google to showcase the competences of its free Android mobile OS. That is setting a standard for other smartphone makers.


There’s a whole heap of interest to realize if Google could possibly create a more competitive chip. That could distinguish its products. Then again don’t let that interest trick us into intelligent that Tensor is closely correspondent to Apple’s A-Series chips. Tensor is the method on a chip. That is by a mix of components that Google itself has designed and others that it has licensed.

What is (TPU) Tensor Processing Unit?

Tensor Processing Unit (TPU) is an AI accelerator application specific integrated circuit (ASIC). It is technologically advanced by Google definitely for neural network machine learning. It is primarily using Google’s own TensorFlow software. Google activated using TPUs inside in 2015.They made them available for third party use in 2018. Together as fragment of its cloud infrastructure and by proposing a minor version of the chip for sale.

Google has correspondingly used TPUs for Google Street View text processing. This was intelligent to discover all the text in the Street View database in less than five days. A separate TPU may process over 100 million photos a day in Google Photos. It is similarly used in RankBrain which Google uses to deliver search results.

It is designed for a high volume of low accuracy computation associated to a graphics processing unit. That is with more input or output operations per joule. And deficiencies hardware for rasterisation or texture mapping. The TPU ASICs are riding in a heat sink assembly. That assembly may appropriate in a hard drive slot within a data center rack. For different types of machine learning models suited different kind of processors. TPUs are healthy matched for CNNs though GPUs have paybacks for some fully-connected neural networks. The CPUs may have benefits for RNNs.

Google delivers third parties access to TPUs. It provide this service over its Cloud TPU service by way of part of the Google Cloud Platform. Also give through its notebook-based services Kaggle and Colaboratory.

Cloud TPU programming model

Cloud TPUs are in no time at performing dense vector and matrix computations. Transferring data between Cloud TPU and host memory is slow compared to the speed of computation—the speed of the PCIe bus is far  slower than both the Cloud TPU interconnect  and therefore the  on-chip high bandwidth memory (HBM). Partial compilation of a model, where execution passes back and forth between host and device causes the TPU to be idle most of the time, expecting data to arrive over the PCIe bus. To alleviate this example, the programming model for Cloud TPU is meant to execute much of the training on the TPU—ideally the whole training loop.

Below are some noticeable features of the TPU programming model:

  • All model parameters are reserved in on-chip high bandwidth memory.
  • The price of initiation computations on Cloud TPU is repaid by performing many training steps during a loop.
  • Input training data is flowed to an infeed queue on the Cloud TPU. A program running on Cloud TPU saves batches from these queues during each training step.
  • The TensorFlow server running on the host machine makes data and pre-processes it before infeeding to the Cloud TPU hardware.
  • Data parallelism: Cores on a Cloud TPU perform a uniform program residing in their own respective HBM during a synchronous manner. A discount operation is performed at the top of every neural network step across all the cores.

Edge TPU​

The machine learning runtime used to perform models on the Edge TPU. That is based on TensorFlow Lite. The Edge TPU is just accomplished of accelerating forward-pass operations. That means it’s mainly useful for executing inferences. The Edge TPU similarly only supports 8-bit math. It requires to moreover be trained using the TensorFlow quantization-aware training method.

Machine learning models trained in the cloud ever more need to run inferencing at the edge. That is, on devices that function on the edge of the Internet of Things (IoT). These devices comprise sensors and other smart devices. That collect real-time data, create intelligent verdicts. Finally take action or connect their information to other devices or the cloud.

Google designed the Edge TPU coprocessor to speed up ML inferencing on low-power devices. Only an individual Edge TPU may execute 4 trillion operations per second. That is doing by using only 2 watts of power. For instance, the Edge TPU may implement state-of-the-art mobile vision models.

Future of ambient computing

The Pixel 6 hardware and software mix rises up the smartphone’s ability. That is due to comprehend what people say in another step toward a future of ambient computing. Google’s transferal to Tensor derives as per the world faces a worldwide chip shortage. That has hopped production of products reaching from cars to computers.

Mansoor Ahmed

Mansoor Ahmed is Chemical Engineer, web developer, a Tech writer currently living in Pakistan. My interests range from technology to web development. I am also interested in programming, writing, and reading.