As the Google was trying to solve the issues related to its voice recognition may result in the formation of AI chips, which can clean the clocks of large chip developers. Almost four years back, Google experienced an issue that if users use the voice recognition services for at least three min daily, Google would be required to make the number of data centers double for handling the numbers.
According to the PC World, Google decided to do some research for making the specialized hardware for driving the machine learning apps such as voice recognition. And after this Tensor Processing Unit (TPU) was made which is a chip that is modeled to speed up the initial levels of profound neutral networks.
According to research done by Google a few days back, the chip is much better than CPU and GPU, in terms of performance and raw power per watt of the total power used. A TPU is 15 to 30 times quicker at performing the machine learning tasks as compared to Intel Haswell CPU or Nvidia K80 GPU. Particularly, the efficiency of the chip for each watt is 25 to 80 times better than that of CPU and GPU.
Google has utilized TPUs in its data centers and has put implied them for improving the performance of the apps such as translation and image identification. The TPUs are majorly advantageous when you consider the energy consumption, which is an important thing linked with the price of hardware at the bigger level.
TPUs are the app specific integrated circuits (ASIC) which show that you cannot use them for flexible purposes as in computers. It is much like a chip which is made to run only for one special purpose. The TPUs are much better at latency than other silicon. Google said that the machine learning techniques require a faster response for providing the good experience for the user.
Google evaluated the chips of six various networks inferences apps, exhibiting 95 % of all these apps in the data centers of Google. After this, using this technique when it comes to the GDDR5 memory which is a part of Nvidia K80 GPU. Google’s research shows that the bandwidth of memory can restrict the performance of few apps.
The company’s research claims that there is a space for additional optimization of software for increasing the performance level, which involves the convolution of the neutral network app as a participant.