Google’s TPU 3.0 Machine Hardware Accelerates and Becomes Faster

The general trend of artificial intelligence is strongly supported by Google’s last developments, as its CEO claims. 

Google is now building up the Tensor Processing Unit (TPU), which is aimed to have a wider range of processing tasks. Since 2016, it has been announced that these units have been used only within the company. However, the follow up is announced to be coming with some major improvements for the new market needs coverage. The main purpose is to advance machine learning and model training. The newest TPU will be primarily applied via Google Cloud service. 

Overall, the third edition has one major improvement, such as water cooling and the new system architecture. The improved 3.0 pods TPUs are said to have more powerful performance. Thus, its power now requires cooling capacities, which are based on copper-plated application systems. 

Besides that the racks are placed tightly unlike the previous edition, they also contain space for the water cooling, presumably supplemented separately with the water pump for that. TPUs are containing in the clusters of up to 64 boards. It is claimed that the last edition has 4 times better performance than the previous one with the HBM storage of 4TB totally and 2,400GB/s memory bandwidth. For the high demanding machine learning, it was apparently a challenge for Google. 

Google understands the machine learning market needs and attempted to shift from computing and advertising primarily benefiting its leadership role in the field. Developing this demanding task is aimed to open up the new capacities for the business, such as learning the human behavior to operate with the bigger data amounts, in addition to the existing advertising market.