This article is more than 1 year old

Nvidia chases defense, intelligence ISVs with GPUs

Chewing through signals and video in real-time

The defense industry already accounts for somewhere between 10 and 15 per cent of Nvidia's Tesla GPU coprocessor sales and is a big chunk of the sales of Quadro high-end graphics cards, as well. And Sumit Gupta, general manager of the Tesla Accelerated Computing business unit at the company, says that there is a lot more business to be had in the geospatial intelligence market – GeoInt in defensespeak –­ and the company will be working with partners to chase those dollars.

Rather than build its own workstations or server appliances, as it has done in creating its GPU accelerated appliance for virtual workstations, called the Visual Computing Appliance announced in March of this year, Nvidia is going after the GeoInt market with partners, and is making sure that the GPU acceleration, math and other function libraries, and applications all line up to make its Tesla and Quadro products the default accelerators for processing audio and static and video image data.

The military and intelligence communities are gathering up huge amounts of information from surveillance, but the image-processing capabilities of systems have been limited by the compute capacity of CPUs and the budgets allotted to build systems and buy the application software to do face and voice recognition from the mountains of data gathered up. At the moment, the processing of this voice and image data is still processed in large part by employees.

To make his point, Gupta says that the US Air Force expects for a whopping one-third of its active duty personnel to be involved in video analysis for the massive piles of video surveillance collected from planes around the world and satellites above it. And to drive it home a little more, Gupta reminded everyone that with the Boston Marathon bombers earlier this year, the Boston police department and the FBI had video of the perpetrators of the crime and both men involved had Massachusetts drivers licenses, but the facial recognition software was too slow and not sophisticated enough to take the images from the video and match them up to the database of license photos.

"Imaging and pixel processing is just a slam dunk for us," says Gupta with a laugh. But, of course, that is only true if the software developers in the GeoInt space have the tools to make their applications run on hybrid ceepie-geepie machines instead of on CPUs alone.

"The libraries are crucial," says Gupta, and Nvidia is not thinking of building its own iron as it has for the Visual Computing Appliance. "There is no single solution that we could build that would fit all needs. And, more importantly, the GeoInt market is full of small system integrators and software suppliers that build the solutions for the defense and intelligence communities. This is more of a platform play than an appliance, which would not play well for these reasons."

Nvidia's GeoInt software partners and the libraries they need

Nvidia's GeoInt software partners and the libraries they need

Given this, Nvidia has put together a set of libraries to make its Quadro graphics cards and Tesla GPU coprocessors do well in the routines and algorithms that are required for modern signal and image processing, and is working with GeoInt ISVs and their system integrators to get applications accelerated by GPUs. They are:

  • Accelereyes ArrayFire: Image and signal-processing functions accelerated by GPUs
  • DelCross Savant: High-frequency ray tracing engine for antenna modeling
  • MATLAB Image Processing Toolbox: Algorithms, functions, and applications for image processing and algorithm development in MATLAB
  • Nvidia CUDA FFT (cuFFT): For signal processing applications
  • Nvidia OptiX: Programmable ray-tracing framework to model stealth designs
  • Nvidia Performance Primitives (NPP): Basic image-processing kernels and functions
  • OpCoast SNEAK: radio-frequency propagation analysis, ray tracing, and jamming analysis development kit
  • OpenCV: Computer-vision functions

To illustrate how GPU acceleration can help defense applications, Gupta trots out the example of Luciad, which creates mission-planning software for the military that does one very important but very difficult task.

Luciad is accelerating its mission planning apps with GPUs

Luciad is accelerating its mission planning apps with GPUs

Luciad's Lightspeed software takes in video data from drones, image data from satellites, and signal data from local reconnaissance, mashes it up with the 3D maps generated by geographical information systems, and provides real-time line of sight mapping software. In plain American, what this means is that Luciad shows you on a 3D map of the area where a mission is going down precisely where the enemy can see you.

When you are in a helicopter mission working for NATO in Afghanistan or Iraq, you want to know this, which is why NATO uses this software. And by making Lightspeed do calculations on GPUs, the line-of-sight analysis can be done in real time, so that as conditions change, the helicopters stay in the green zones in the map above, where they can't be seen, and stay out of the red zone, where they can be spied.

Lightspeed is not just used by military organizations around the world, but also by the US Federal Aviation Administration to track airplanes, and getting better real-time tracking is a goal.

The speed-up for the algorithms used in signal and video processing can be significant when dumping the work from CPUs off to GPUs, as this benchmark data shows:

The algorithm libraries commonly used in GeoInt processing see a 5X to 10X speedup on GPUs

The algorithm libraries commonly used in GeoInt processing see a 5X to 10X speedup on GPUs

Nvidia is not prescribing any particular hardware for GeoInt applications, but is suggesting that dual-socket Xeon E5 workstations be equipped with Tesla K20 GPU coprocessors for number-crunching and Quadro 4000 graphics cards for displaying 2D and 3D applications. For back-end servers, a base GeoInt setup recommended by Nvidia is four two-socket Xeon E5 servers, each equipped with two Tesla K20 coprocessors. ®

More about


Send us news

Other stories you might like