GPUs shift the computing paradigm: A 10 to 100x performance increase coming soon to a military system near you: Interview with Kevin Berce, Business Development Manager at NVIDIA
Why is real-time video rendering about to get a whole lot better? NVIDIA's CUDA programming environment, explains Business Development Manager Kevin Berce.
GPU manufacturer NVIDIA has not only been keeping up with the times by offering processors for smartphones and tablets; it has also been enabling key shifts in the defense paradigm: Remember the days when it took four to six hours just to render one hour of UAV video? Now NVIDIA GPUs enable UAV video rendering in real-time. And NVIDIA is set to deliver an ultra-accelerated, GPU-enabled 10 to 100x performance increase for the defense industry, as a recent interview with Kevin Berce, NVIDIA Business Development Manager, reveals. Edited excerpts follow.
MIL EMBEDDED: Let’s start with a high-level overview of NVIDIA.
MIL EMBEDDED: OK, so which types of technology is NVIDIA focusing on these days?
MIL EMBEDDED: Which brands fall under which market segments then?
MIL EMBEDDED: So remind us about GPU computing.
MIL EMBEDDED: Let’s drill down on the Tesla group for a moment. What’s involved there?
MIL EMBEDDED: Do you have any numbers on how fast Tesla runs in supercomputers?
BERCE: ], Tesla GPUs are powering three of the top five supercomputing systems in the world. Two of these systems are in China and one is in Japan.
The systems are dealing with big computational challenges. It’s also worth noting that the Tsubame system is the world’s Greenest Production Supercomputer, meaning that it has Petaflop-class performance; but it’s also extremely power efficient, only consuming slightly over 1 megawatt. This characteristic is something uniquely enabled by GPUs.
MIL EMBEDDED: What is the biggest power consumption? Tens of megawatts?
BERCE: GPU-accelerated supercomputers require about half the power compared to a CPU-only supercomputer. NVIDIA strives to reduce power requirements with GPUs, thus I do not know what the largest power consumption is with other CPU-only systems.
MIL EMBEDDED: What value are GPUs adding to the defense industry these days?
BERCE: There are basically six verticals inside the defense space that we’re focused on, where the GPU is adding value of anywhere from a 10 to 100 times performance increase: satellite imaging, video enhancement, aerodynamics/CFD, computer vision, signal processing, and electromagnetics. Our defense GPU customers include system integrators, defense contractors, and many other partners.
MIL EMBEDDED: So what types of technical problems can NVIDIA GPUs solve, say, in UAVs or video systems?
BERCE: One of our customers, Motion DSP, for example, has software called Ikena ISR. The problem they’re addressing is that when video sensors get deployed in platforms like UAVs, the video acquired is of a challenging quality to obtain intelligence from.
MIL EMBEDDED: Tell us a little about the hardware that’s being used with the Ikena ISR software.
BERCE: It’s actually just a Dell workstation and inside the Dell system [running Windows], there is a CPU and one Tesla GPU, the C2070. The C2070 has 448 GPU cores, 6 GB of memory, and uses a PCI Express card that is passively or actively cooled. The value the GPU brings to the workflow is the ability to process multiple tasks in real time. Without having a GPU in the mix, you wouldn’t be able to do this in real time.
MIL EMBEDDED: You’d have to render video offline and then get it all cleaned up.
BERCE: Exactly, in the Motion DSP example, if the system didn’t include a GPU, an hour of video would take somewhere between four and six hours to render. But with the GPU, it’s rendered immediately.
MIL EMBEDDED: What about ISR video analysis powered by a GPU?
BERCE: IntuVision makes a software product called Panoptes, which is designed to enable analytics on streams of video coming in. Often in military reconnaissance or for high-end retail establishments, you have video surveillance at many different spots you want to monitor and might only have one person monitoring up to 25 video feeds. It’s kind of challenging because at any given point, you can have human error and just miss something. But, having that process automated and offloaded to a computer, you know it’s going to sift through all that data. With a CPU, you can roughly process 4 HD streams at 3-5 frames per second, but we’re told that to get really valuable intelligence out of it, processing up to 20 frames per second is needed. By adding a GPU into this mix, you get a 12x speedup for better object tracking and real-time notification for up to 90 HD streams.
MIL EMBEDDED: Do most vendors buying your products – whether they’re software box companies or embedded guys like GE – use a single GPU instead of multiple GPUs, typically, then?
BERCE: The answer depends on the application they’re using. They can use up to eight GPUs in a single operating system. We’re seeing more and more use cases where folks are using more than one GPU. A challenge DARPA has is the desire to see how much a simulation would speed up when you add additional GPUs to the mix. That’s because one of the grand challenge problems they have in computational fluid dynamics is that a single 30-second simulation takes 150,000 CPU hours. So if they can make that simulation GPU aware, they can reduce the 150,000 CPU hours substantially.
MIL EMBEDDED: Switching gears, rocks are being thrown at NVIDIA because of the closed-naturedness of CUDA versus the freely licensed, open source OpenCL environment.
BERCE: The CUDA programming environment was essentially what enabled NVIDIA to turn the GPU from being a regular graphics processor into a massively parallel processor that can handle the type of work that we’ve been talking about today. It was a far cry from the world of traditional GPGPUs. If you remember about six or seven years ago, people talked about using GPUs for computing, but they were having to translate their code from a computational language into a graphics language so that the GPU could understand it. It was very difficult. CUDA enabled developers to write in industry-standard languages to access the GPU, so that the very first version was released with CUDA C; and it was basically very similar to C. You just added a few keywords and additional instructions to change your algorithm to understand targeting a many-core processor versus fewer cores on a CPU. So there’s CUDA C and now there is also CUDA C++. Third parties such as the Portland Group offer CUDA FORTAN as well. Most importantly, last week, the Portland Group introduced CUDA x86, which enables developers to compile CUDA codes to run on CPUs instead of GPUs.
So I wouldn’t say that it’s fair to call CUDA “closed” at this point – it supports multiple languages and open standards and can be modified to run on other architectures. CUDA will continue to be our platform for innovation; we have the ability to innovate incredibly fast. And we are able to add new features very quickly, which is something that customers in this space are very keen on us continuing.
NVIDIA 408-486-2000 www.nvidia.com