AI and deep learning have enabled unprecedented innovation in every industry, from healthcare to manufacturing to retail and more. As the AI continues to power the next industrial revolution, developers and innovators are challenging themselves to stay ahead of the game by seeking more and more computational power. In the past, the limiting factor for AI training was a simple matter of GPU accessibility. Today, the complexity of deep learning models and the size of datasets presents new challenges, now addressed by NVIDIA DGX-2, the newest member of the NVIDIA DGX family of AI supercomputers.
DGX-2 represents an architectural landmark, integrating 16 Tesla V100 GPUs using a revolutionary internal “superhighway” network of GPU interconnectivity that takes deep learning for enterprise to an unprecedented level.
As NVIDIA’s Principal Architect Glenn Dearth puts it, DGX-2 is a work of art, science fiction come to life. “When I look at that interconnect, it’s magical,” he says at GTC 2018. The magic he’s referring to is drawn from interweaving NVIDIA DGX software with a brand new interconnect architecture built on NVIDIA NVSwitch, the backbone of the world’s first 2 petaFLOPS GPU accelerator.The magic he’s referring to is drawn from interweaving NVIDIA DGX software with a brand new interconnect architecture built on NVIDIA NVSwitch, the backbone of the world’s first 2 petaFLOPS GPU accelerator.
NVSwitch, delivers an ultra-high speed, low-latency network fabric for GPUs, eliminating bottlenecks associated with traditional PCIe based approaches, enabling the 16 Tesla V100 GPUs to connect in a unified memory space, united as one. This capability will allow data scientists and AI innovators to unleash their full creativity, building the most complex deep learning models, now training the previously untrainable, answering the world’s most pressing AI challenges making reality out of science fiction.
Want to get a look under the hood of the world’s first 2 petaFLOPS AI supercomputer? Watch Glenn Dearth and Vyas Venkataraman present Inside DGX-2 at GTC 2018 here.