Obsessed with technology?
Subscribe to the latest tech news as well as exciting promotions from us and our partners!
By subscribing, you indicate that you have read & understood the SPH's Privacy Policy and PDPA Statement.
News
News Categories

Meet the US$129,000 NVIDIA DGX-1 deep learning supercomputer

By Vijay Anand - on 6 Apr 2016, 2:31pm

Meet the US$129,000 NVIDIA DGX-1 deep learning supercomputer

Unlike the DIGITS DevBox deskside deep learning machine, the DGX-1 is rack based deep learning supercomputer - a first of its kind.

When you think about the term “supercomputers”, you’re likely to think about massive office floor sized rooms lined with super cooled computers end-to-end. That still holds true to this day. So did NVIDIA finally release its own supercomputer to challenge the likes of Oak Ridge National Laboratory and TianJin National Supercomputer Center

Far from it, but the NVIDIA DGX-1 is a supercomputer in relative terms when compared to other equivalent form factor systems. The NVIDIA DGX-1 is a rack size sized system that is touted to provide processing throughput that’s equivalent to 250 typical dual-Xeon based rack servers!

A typical dual Xeon server rack compared against the new NVIDIA DGX-1 'supercomputer'.

NVIDIA's equation: 1 x DGX-1 = 250 traditional server boxes.

What’s under the hood of the NVIDIA DGX-1?

This is what’s underneath the hood of the NVIDIA DGX-1. Click to see enlarged image.

In a nutshell, the NVIDIA DGX-1 is the world’s first purpose-built high performance system for deep learning with all the necessary hardware components and software layers in place for quick deployment. If you recall, last year at GTC 2015, NVIDIA debuted the DIGITS DevBox deskside deep learning machine? Well that still only ran four GeForce GTX Titan X (Maxwell class) GPUs. This presentation slide shows how the best deep learning box of 2015 compares with that of 2016:-

The NVIDIA DGX-1’s awesome processing throughput it possible only because it uses 8 of the very latest Tesla P100 Pascal-based GPUs. That plus Pascal’s new architectural highlights for massive chip-to-chip intercommunication boost through NVLink and the use of HBM2 memory really notches up the entire system’s capability. Here’s what else you can expect out of the DGX-1 ‘supercomputer’:-

  • Dual 16-core Intel Xeon E5-2698 v3 (2.3GHz) processors
  • 512GB of 2133MHz DDR4 memory (LRDIMM)
  • NVLink Hybrid Cube Mesh
  • 8 x Tesla P100 GPUs with 16GB of HBM2 memory each
  • A grand total of 28672 CUDA cores
  • Total processing throughput:- 170TFLOPS (GPU FP16) and 3TFLOPS (CPU FP32)
  • 4 x 1.92TB SSD RAID 0
  • Dual 10GbE + Quad Infiniband 100GbE
  • 3200W PSU
  • 3U Rack form factor
  • Software: Ubuntu Server Linux OS, DGX-1 recommended GPU driver, NVIDIA Deep Learning SDK, etc.

The NVIDIA DGX-1 software stack includes major deep learning frameworks, the NVIDIA Deep Learning SDK, the DIGITS GPU training system, drivers, and CUDA, for designing the most accurate deep neural networks (DNN). This powerful system includes access to cloud management services for container creation and deployment, system updates, and an application repository. The combination of these software capabilities running on Pacal-powered Tesla GPUs allows applications to run up to 12x faster than any previous GPU-accelerated solutions (as shown in the photos above).

Price and availability?

With all of the hardware it boasts, it’s no wonder why NVIDIA slaps an US$129,000 list price for the DGX-1. In fact, this purpose built server that’s designed for deep learning costs much more than US$50,000 Quadro VCA server appliance.

The DGX-1 is expected to be available from June this year and is designed for compatibility with existing data centers.

Who should get the NVIDIA DGX-1?

Basically anyone who needs to supercharge deep learning performance. A higher performance training accelerates productivity, which in turn delivers speedier turnaround of insights for faster innovation and or other critical analysis.

Chiefly, the DGX-1 is marketed at data scientist and AI researchers who require accuracy, simplicity and speed of deep learning success.

For a start, NVIDIA will be closely working with the following pioneers of AI research to help deliver the DGX-1 to those who need it the most and for whom the box was made for. After that, NVIDIA will be processing order requests submitted on their website.

Loading...