Nvidia Tesla V100: 120 TFLOPs with 5376 CUDA nuclei


Nvidia have presented the first chip of the new Tesla generation, which offers a huge performance jump to the Pascal generation.

As last year, Nvidia is using the spring to introduce a new graphics card generation. The new generation of Tesla will probably be launched in the fall, the first information on the new generation, as well as a taste for the professional accelerators, Nvidia has already presented at the in-house conference. The chip area of ​​the first Volta GPU will be 815 mm², the largest chip area for a GPU ever officially presented.

The first model with the chip will not be a consumer graphics card, but a graphics accelerator for servers in the professional field. The Volta GPU, named GV100, will have six Graphics Processing Clusters, each with 14 Volta Streaming Multiprocessors. The number of CUDA cores increases from 3,584 in the GP100, the Pascal generation, to 5,376 of the new Volta generation. The theoretical computing power of the chip increases to 120 TFlops, with the deep learning performance. The singlePrecision performance is specified with 14.9 TFlops.

The new Volta GPUs are manufactured in TSMC’s 12 nm design. The GV100 chip can boast 21.1 billion transistors. The new graphics accelerator provides 16 gigabytes of HBM2 memory, which is connected to the system with 4,096 bits. The GPU is to be operated with 1.455 MHz, while the memory with 900 MHz goes to work. The maximum power dissipation of the graphics accelerator will not exceed 300 watts. The Tesla V100 will be delivered in the third quarter of 2017.

Share

Leave a Comment