An Nvidia's seventh annual GPU Tech Conference, chief executive Jen-Hsun Huang had some impressive new hardware to demo: Nvidia's DGX-1, a GPU-driven server that the company hopes will find a home in everything from research institutions to mid-market businesses looking to gain a competitive edge in deep learning.
While the deep learning land rush is driving the need for more storage, Nvidia is hoping its new hardware can help businesses outside of Silicon Valley actually process and learn from all new newly created and collected data they're generating.
And to that end, the device, shipping in Q2 of this year, has some impressive specs: 170 teraflops thanks to eight of Nvidia's brand new Tesla GP100 GPUs, as well as two 16-core Intel Xeon E5-2698 v3 2.3GHz CPUs, and 512GB of DDR4 RAM.
It also features integration with a variety of deep learning frameworks, including Nvidia's own offering. What kind of difference can the $129,000 make? Nvidia said the DGX-1 could crunch one sample training set in two hours, compared to 6.25 days days with a typical dual socket CPU. That means much faster iteration and more room for experimentation.