NVIDIA's DGX-1
According to NVIDIA, the 132 lb (60 kg) DGX-1 has a 170-teraflop performance and is tuned to focus on solving problems that are still mystifying engineers working on deep learning and artificial intelligence (AI).
To power its new supercomputer-in-miniature, NVIDIA has packed the DGX-1 with eight 16 GB Pascal-powered Telsa GP100 GPUs, 512 GB of RAM, and a 1.92 terabyte solid state drive. What’s most interesting about about the DGX-1 is that NVIDIA has tied together each of the machine’s GPUs via its Pascal architecture and NVIDIA NVLink, a high-speed bi-directional chip interconnect. With these two elements in place, applications developed for the DGX-1 can be processed much faster, reducing the machine’s time to solution. Add in the machine’s form factor and the “plug-and-play” nature of its design, and it seems like NVIDIA might have actually developed something very interesting.
But what’s in it for the engineer working outside of computer science’s bleeding edge?
While NVIDIA has positioned the DGX-1 as a machine determined on breaking down machine learning problems, it’s interesting to wonder what a machine like the DGX-1 could do for engineering. If the right application program interfaces were in place, supercomputers could be used to run all manner of aerospace optimization algorithms, interpolate nuclear reaction data to advise designers on how to build better reactors, or even suggest better product lifecycle management options for large scale projects. Really, the sky’s the limit, and I don’t imagine that NVIDIA, or any other hardware manufacturer interested in building smaller, more easily configurable supercomputers to stop at the AI research market. It’s very possible that that DGX-1 is just the first in a new wave of supercomputers that will be more accessible and appealing to business.
Though NVIDIA has let the world know that the DGX-1 is a thing, the company hasn’t set a release date for its supercomputing dynamo. That being said, NVIDIA was kind enough to let everyone in on its retail price. For a whopping $129,000, you and your team can have your very own supercomputer.