Recently, at the GTC 2017 conference, NVIDIA officially released the most powerful nuclear bomb to date - the flagship computing card Tesla V100.

Tesla V100 is a product based on Volta architecture, with 5120 CUDA units built in, core frequency of 1455MHz, 16GB HBM2 video memory, single precision floating-point performance of 15 TFLOPS, double precision floating-point performance of 7.5 TFLOPS, and video memory bandwidth of 900GB/s.

In addition, the Tesla V100 also adds a Tensor unit highly related to deep learning. The performance of the Tensor is claimed to be 120 TFLOPS. This has enabled the Tesla V100 to have more than 21 billion transistors, with an unprecedented core area of 815 square millimeters, and is manufactured using the 12nmFFN process.

By comparison, the last generation flagship Tesla P100 has only 3584 CUDA units, with 15.3 billion transistors and a core area of 610 square millimeters. In addition, the L2 cache of the Tesla V100 has increased from 4096KB of the Tesla P100 to 6144KB, and the total register file size of each group of SM units has increased from 14336KB to 20480KB.

NVIDIA said that the Tesla V100 will first be used for deep learning and supercomputing DGX-1. There are eight Tesla V100 computing cards inside, and the peak computing performance is as high as 960 TFLOPS. It is said that the workload of TITAN X in eight days can be completed in eight hours.

The price of DGX-1 is 149000 dollars（ About 1.02 million yuan ）Why is it so expensive? Because this is not a civilian graphics card, but don't worry, it won't be long before this civilian class will come out, and from the number of built-in CUDA units, the performance will be very good.

Follow our Weibo @ Love Computer

Follow our WeChat official account: playphone

Of course, we also pay attention to our Bilibili account: love computer