Google announces 4th generation processor 'TPU v4' specialized in machine learning, more than double the performance of the previous generation



The 4th generation model ' TPU v4

' of the dedicated processor 'Tensor Processing Unit (TPU) ' developed by Google, which is specialized for machine learning, was announced at the developer conference 'Google I / O 2021' on May 18, 2021. it was done. The new processor has already been deployed in Google's data center and will be available to Google Cloud users in late 2021.

Google unveils 4th generation Tensor Processing Unit --9to5Google
https://9to5google.com/2021/05/18/google-unveils-4th-generation-tensor-processing-unit/

Google details new AI accelerator chips | VentureBeat
https://venturebeat.com/2021/05/18/google-details-new-ai-accelerator-chips/



In recent years, AI has played an important role in various fields, but since a huge amount of arithmetic processing is required to train a large amount of data, it is important to use a powerful machine to increase productivity. is. Google is optimized processor in TPU machine learning and deep learning that will develop, 2016 to place the Google of artificial intelligence ' AlphaGo ' to be the top players of Go Lee Cedres (Riyo乭) of Kudan opposite station but It is said that TPU was used.

Google is continuously introducing a new generation model of TPU, and in 2018, the third generation model 'TPU v3' was announced.

Google announces the 3rd generation processor 'TPU3.0' dedicated to machine learning, in a situation where cooling can not catch up and even a liquid cooling system is introduced --GIGAZINE



And on May 18, 2021, Google announced the 4th generation model of TPU, 'TPU v4'. Compared to the previous generation model 'TPU v3', TPU v4 has more than doubled the index, FLOPS, which indicates how many floating-point operations can be performed per second.

Further, the memory bandwidth by the benefit of interconnect technology has also greatly improved, unit called 'pod' that combines 4096 units TPU v4 is 1ExaFLOPS ( Exa can provide operation performance FLOPS) ones. This is equivalent to a notebook PC processor that has reached a peak performance of 10 million units.

Google CEO Sundar Pichai said, 'This is a historic milestone for us. Previously, we had to build a custom supercomputer to get the computing power of exaFLOPS, but we Is already deploying many TPU v4s. Soon there will be dozens of TPU v4 pods in the data center, many of which are powered by energy that does not emit nearly 90% of carbon. ' In late 2021, Google Cloud users will be able to use the TPU v4 pod.



in Hardware, Posted by log1h_ik