Google releases third generation AI chip TPU 3 instead of NVIDIA GPU

category:Internet
 Google releases third generation AI chip TPU 3 instead of NVIDIA GPU


Figure: Google CEO Sundar Pichai (SundarPichai) introduces the companys third generation AI chip at the developer conference. In addition, if the new version of TPU is similar to its predecessor, it will also be opened to third party developers through Googles public cloud service, which will help Google compete with Amazon and Microsoft. Earlier this week, Microsoft released a special chip that could be used in its Azure cloud platform. Pichai said that when people use the third generation of TPU on a large scale, they can enjoy more powerful computing power. Now, he said, the performance of each chip is 8 times that of last year, far more than the 100Petaflops (Petaflops: per second one thousand trillion / ten million (10^15) floating-point operation). At present, the 16 boxes of the latest GPU of NVIDIA can only provide 2Petaflops computing power. Last years version of TPU has shown good results. The test results published in recent months show that in some cases, the second generation of TPU can provide better performance than the existing GPU, although TPU does have some limitations, such as lack of support for the FacebookPyTorchAI software framework. PyTorch open source community is always trying to change this situation. As early as 2013, Google began to envisage the concept of TPU, a chip specially designed to support AI. Since 2015, Google has started using TPU chips internally to support GooglePhotos, Google search results or GoogleCloudVisionAPI. Google first publicly acknowledged the existence of TPU in 2016, and the second generation of TPU was launched last year at the developer conference. (small) source of this article: NetEase science and technology report editor: Wang Fengzhi _NT2541 Last years version of TPU has shown good results. The test results published in recent months show that in some cases, the second generation of TPU can provide better performance than the existing GPU, although TPU does have some limitations, such as lack of support for the FacebookPyTorchAI software framework. PyTorch open source community is always trying to change this situation. As early as 2013, Google began to envisage the concept of TPU, a chip specially designed to support AI. Since 2015, Google has started using TPU chips internally to support GooglePhotos, Google search results or GoogleCloudVisionAPI. Google first publicly acknowledged the existence of TPU in 2016, and the second generation of TPU was launched last year at the developer conference. (small)