How NVIDIA Built a Supercomputer for the Office

NVIDIA DGX Station brings together hardware, software and industrial design to provide dramatic boosts to deep learning speeds.
by Markus Weber

Supercomputing at the deskside is making inroads into the enterprise.

Enterprise users often need the extreme performance of a data center, but without the noise of air conditioning equipment in overdrive and the mess of cabling and other equipment.

That’s precisely what the NVIDIA DGX Station was designed to do. It’s a deep learning supercomputer designed for the office, yet provides twice the performance of the most powerful workstations available.

“DGX Station is very quiet in the office thanks to its liquid cooling. And it’s superfast — it boosted the speed of our deep learning training by 170 times,” said Frank Wu, head of the Machine Learning Business Network at SAP Leonardo Machine Learning.

Chris Klein, lead architect for DGX Station at NVIDIA, was tasked with building the system. The goal: design a supercomputer for the deskside that could draw power from a normal wall socket, use plug-and-play GPU-accelerated software and operate not much louder than a whisper.

Crafting DGX Station required enlisting NVIDIA engineers across a number of disciplines to work together on this purpose-built, one-of-a-kind machine for deep learning.

“This could not have been accomplished at a normal company with work silos. Everybody had to be committed,” said Klein. “An undertaking like this had never been done before.”

Making of the DGX Station

Planning such highly capable supercomputing, thermal work in copper plumbing, mechanical engineering and an elegant, compact industrial design would be no small feat.

Supercomputer specs had to be squeezed into an ambitiously small desktop-size box. Inside, it’s packing four Tesla V100 GPUs, NVLink interconnect technology, 128GB of GPU memory and 20,480 NVIDIA CUDA cores — all helping it deliver 500 TFLOPS of deep learning punch.

“Our automated fashion image tagging tool was trained on DGX Station and we experienced 100x faster than human tagging, 90 percent operation costs reduction and higher accuracy than human tagging,” said Jaeyoung Jun, CEO at Omnious. “This is helping us stay ahead of the fashion trend and provide timely service to our customers.”

DGX Station provides its enormous computing heft inside the office, which requires it to be water cooled — or risk being as loud as a running dishwasher. The industry standard for air-cooled workstations is about 45 decibels. DGX Station bests that at 35 decibels, the volume of typical office ventilation systems.

Pleasant office manners aside, DGX Station is a deep learning speed demon: It’s 2x faster than the zippiest workstations. And the sleek, svelte box takes the place of four server racks while sipping one-twentieth the energy.

“With the NVIDIA DGX Station, we see a 3x reduction in training time, while maintaining over 90 percent accuracy in our real-time, multi-sensor event detection and classification solutions for telco, IoT, security and Smart Cities,” said David Ohm, co-founder and CEO of KickView, an AI systems maker.

“With the flexibility and unparalleled performance provided by DGX Station, we can finally train and apply state-of-art deep learning models with minimal overhead, delivering a real-time AI solution for faster and safer medical imaging exams,” said Enhao Gong, CEO and co-founder at Subtle Medical.

DGX Station doesn’t stop at obsessively engineered hardware. We’ve paired it with our widely used deep learning software stack, which ports to DGX-1 and gets monthly updates from our deep learning and software engineering teams, so performance gets better over time. Users can power it on and start working with their data.

The result: NVIDIA DGX Station is ready for business.

For a limited time, buy four DGX Stations and get one free