Nvidia today unveiled a professional graphics card Quadro GV100. This is the first adapter in the family based on GPU GV100 generation Volta. Now we have three cards of the Volta: V100 Tesla for supercomputers, Quadro GV100 for developers and Titan V for gamers, for professionals.
Quadro GV100 got 5120 active CUDA cores as the Titan V and Tesla V100, whereas in General, the GPU them 5376 pieces. Active and 640 tensor Tensor processor Cores, designed for operation of systems of artificial intelligence, machine learning and deep learning.
The product has 32 GB of memory HBM2 with a bus width of 4096 bits with the capacity of 870 GB/s. Performance declared equal to 14.8 TFLOPS (FP32), and the performance of tensor block — 118,5 TFLOPS, but this is only for specialized tasks, that is, compared with normal performance impossible.
In contrast to the model of the Titan V, Quadro GV100 can be combined in bundles of two boards, using proprietary NVLink interface with a bandwidth of 200 GB/s. one can Also note a TDP of 250 watts, the presence of two connectors capitania (8+6) and four DisplayPort 1.4. Buy an adapter will be available in April, after parting with 9000 dollars.
Separately, Nvidia said hardware support for the RTX, which we a few days ago, we talked in more detail. Sorry, details of the company are not disclosed. Given that for other generations of hardware technology support ray tracing is not stated, we can assume that it at least partially rests on the same block with a tensor kernel, which simply did not exist in previous Nvidia GPUs.