Nvidia has presented its own supercomputer at the ISC High Performance in Frankfurt Main event, which immediately climbed into 22nd place in the world ranking list. Even more interesting than the performance, however, is the approach behind the DGX SuperPOD: It links 96 DGX-2 already presented a year ago.
Like its 96 individual components, each set to 16 Volta GPUs ( Tesla V100 32 GB ), so the DGX SuperPOD is primarily intended for the AI calculation, explains Nvidia.
Within a DGX-2, the 81,920 shaders of the 16 GPUs are connected via the NVSwitch, which itself includes 2 billion transistors. Each DGX-2 uses an Intel Xeon Platinum 8174 with 24 cores. For the presentation of the cube system, Nvidia called as purchase price around 400,000 US dollar before taxes. The DGX SuperPOD adds up to nearly $40 million.
While turning to the ARM architecture for future exascale supercomputers, the Nvidia group demonstrates the flexibility of its GPU acceleration solutions for artificial intelligence by announcing the arrival of the powerful DGX SuperPOD supercomputer.
With the processing power of 9.4 petaflops, it is positioned as the 22nd most powerful supercomputer in the world but has only required three weeks of implementation, compared to 6 to 9 months for supercomputers of standard architecture. server base.
It is based on 96 Nvidia DGX-2H systems each integrating 16 Nvidia Tesla V100 accelerators (ie 1536 Tesla V100 for the entire system) and interconnected by Mellanox technologies, with the advantage of a modular design.
The primary goal of DGX SuperPOD will be to train the deep neural networks of the artificial intelligence of the future autonomous cars of the market. With 1TB of data generated each day by each vehicle and tracking over several years, these are gigantic amounts of data that need to be processed and analyzed to reduce the potential weaknesses of artificial intelligence on certain scenarios and re-train the networks. neurons to make them more efficient.
Nvidia puts forward its processing capacity, with the example of training the ResNet-50 neural network in less than 2 minutes when it took 25 days in 2015 when it was commissioned with Nvidia K80 GPU processing.