of GPU Technology Conference NVIDIA has announced the GPU GV100 and the accelerator card Tesla V100. They are based on the Volta architecture, which last year introduced architecture Pascal followed by the spring. As it was then, not GPUs for the consumer market in the foreground, but the possibilities for deep learning, high-performance computing (HPC) and business are the first Volta products. The GPU architecture is suitable for both computing and graphics.
the presented by NVIDIA CEO Jen-Hsun Huang flagship Tesla V100 should bring in particular progress of deep learning and artificial intelligence. The graphics accelerator based on the GPU GV100 and impresses by its size. The graphics chip claimed an area of 815 square millimeters and is the largest of a series production. Its production takes over the Taiwanese contract manufacturer TSMC in the 12-nm-FinFET method.
the manufacturer specifies a boost 1455 MHz maximum clock rate and a TDP of 300 Watts. The GPU offers a total 21.2 billion transistors and provides more than 5,000 shader units. The tensor 672 cores are optimized for deep learning. NVIDIA according to these new cores offer the four-time performance compared to its predecessor, Pascal. Thus, in processing unit (TPU) tensor could surpass the performance Google’s special chip which is used to speed up by machine learning for 2015 in Google’s data centers.
“Tesla V100 was designed for the convergence of HPC and AI”, postulated the graphics specialist in a blog to the Volta architecture. It offers a platform for HPC systems to prove himself both computer science and data science and to provide new insights.
“Artificial intelligence drives the greatest technological advances in human history,” said NVIDIA Chief Huang. This will automate intelligence and inspire a wave of social progress, as it was not seen since the industrial revolution. “Deep learning has as a bahnrechene AI approach that ensures learning computer software, an insatiable hunger for computing power. Thousands of NVIDIA developers have worked for three years on Volta and thus helped to meet this requirement – so that the industry can exploit the life-changing potential of artificial intelligence.”
Tesla V100 will come first in NVIDIA’s customized compute servers to use. Eight of the graphics accelerator, see the DGX-1 course, which comes at the price of 150,000 dollars in the third quarter to be delivered. Scaled-down versions of the V100 are also in preparation, including a card that comes out with 250 watt TDP. With reduced clock rate and disabled computing cores, a 150-watt version is also to be expected.