DigitalGlobe and NVIDIA on the Quiet Revolution

astrickBACKGROUND Geospatial technologies, and therefore geospatial solutions (especially those dealing with large imagery files) have always been burdened by high computing challenges. These challenges have translated into latency issues. Over the last several years, DigitalGlobe has invested in GPU technology to accelerate speed to market in delivering imagery products and services. We spoke with Jason Bucholtz about how DigitalGlobe is using GPU technology to improve operations and create a technology infrastructure to support next-gen imagery-based applications. We also spoke with Kevin Berce of NVIDIA, a leader in GPU technology, on the future of advanced computer processing power.


Summary: This interview discusses the operational and customer experience enhancements to satellite imagery products resulting from investments in GPU computing architecture.


LBx: What’s your focus at DigitalGlobe?

JASON:  I’m an architect who deals with data. I’m interested in and responsible for determining how quickly raw data can be turned into information. Imagery that hits the ground from the satellite is “Big Data;” it needs to be processed, and the faster it is processed, the faster organizations have access to the data to answer all sorts of questions about our changing world, and to respond to emergency situations and save lives.

LBx: How do you achieve faster processing?

JASON:  DigitalGlobe invested quite a bit over the last several years in GPU (Graphic Processing Unit) technology. GPUs are high-performance computing processors that accelerate image processing applications to enable more, higher quality and faster work. GPUís enhance traditional high performance computing architectures, by breaking down complex instructions into simpler steps and enabling more simple steps to be executed at the same time. This simplification dramatically increases computational speed. Our goal is to achieve faster processing of pixels. GPU technology has primarily been used in the video gaming sector, and is incredibly well suited to geospatial imagery files.

LBx: Kevin, how would you describe GPUs in a nutshell?

KEVIN: GPUs are being used to power everything from supercomputers to smartphones. Technically speaking, GPUs excel at parallel processing, and that makes them ideal to accelerate image processing and other computationally intensive applications. The goal of employing GPU technology is to reduce the processing time on algorithms (improved customer experience) and reduce the amount of computer rack space, which reduces the sheer number of systems required (reduced operational costs).

LBx: What is driving GPU adoption?

KEVIN:  There are lots of drivers, including the need to increase raw performance of computationally intensive applications, the need to do more processing with less hardware and power, and enabling greater mobility. Big Data and graphically complex video games are driving the need for more horsepower to run analytics and simulations faster. Relative to geospatial applications and the federal government, GPUs are being used to accelerate the processing of Big Data and analytics in the cloud. The issue with Big Data is that the massive volumes of source data can’t be processed quickly enough with typical CPU-only computing solutions.

LBx: What does the faster processing get you?

JASON: It lowers our operating costs for one, and most importantly, it delivers faster analysis to our customers. With GPU computing we can deliver a meaningful geographic image from the tasking of the satellite to delivery to the customer within as little as 120 minutes. That may still seem like a long time in a real-time tweeting world, but that is amazing speed given the complexity of satellite imagery processing workflows. Feature extraction takes more time, but with advanced algorithms we can identify objects on the ground faster ñ for example, cars in parking lots.

GPUs enable us to speed up common preprocessing steps such as orthorectification and pan sharpening. See Figure 1. Orthorectification is the process of tying a pixel to a known location on the earth to render an accurate geographic reference from an image. Pan sharpening merges high-resolution black and white images with lower-precision color bands and fills in the color for a precise high-resolution color image. Intermediate imagery products that include feature extraction and the type of analysis I mentioned earlier all depend on this base-level processing. And thatís the heavy lifting that is performed extremely well by GPUs.

KEVIN: Speed gains depend on the application, but can be as much as 20x or more. Cost savings on the reduced hardware footprint can be as much as 70%.


FIGURE 1. This illustrates the satellite tasking, imaging, downlink, and image processing workflow chain. Graphic courtesy of DigitalGlobe.
FIGURE 1. This illustrates the satellite tasking, imaging, downlink, and image processing workflow chain. Graphic courtesy of DigitalGlobe.

LBx: Can you give some examples of how companies are using satellite imagery?

JASON: Visual audits of financial projections are becoming increasingly popular, for example nancial analysts can use satellite imagery to vet financial projections by analyzing the actual number of cars in the parking lot and estimating earnings per share. Financial analysts can calculate the volume of oil being shipped from satellite imagery. Sounds crazy, but because we know the sun’s position, the height of a cylinder, and the arc of the cylinder’s shadow, we can calculate the volume of the oil in the cylinder. These measurements are now automated and provide financial analysts with critical market information.


imagethumb1

KEVIN BERCE has more than 15 years of experience in high performance computing as director of sales, business development and other customer-facing roles in the United States. Kevin is currently Manager of Business Development for NVIDIA. In his current role, he has responsibility for NVIDIA’s Defense and Intelligence business inside the United States.


Jason bucholtz has been designing and implement- ing large scale high-performance compute systems and large data infrastructures for 15 years. As a consultant to numerous Fortune 500 companies and national labs, he and his teams have integrated over 40,000 HPC nodes and multi-petabyte data systems. At Digital- Globe, Jason designs the next generation compute and storage systems that power the company’s leadership in automated image processing. The pioneering e orts by Bucholtz and team in the area of GPU-based HPC are playing a key role in the company’s business transforma- tion. To date, this transformation has already resulted in a dramatic acceleration of DigitalGlobe’s supply chain, ultimately enabling the company to provide maximum value to customers in the form of both high-definition imagery and geospatial insight.

JASON BUCHOLTZ has been designing and implementing large scale high-performance compute systems and large data infrastructures for 15 years. As a consultant to numerous Fortune 500 companies and national labs, he and his teams have integrated over 40,000 HPC nodes and multi-petabyte data systems. At Digital- Globe, Jason designs the next generation compute and storage systems that power the company’s leadership in automated image processing. The pioneering e orts by Bucholtz and team in the area of GPU-based HPC are playing a key role in the company’s business transformation. To date, this transformation has already resulted in a dramatic acceleration of DigitalGlobe’s supply chain, ultimately enabling the company to provide maximum value to customers in the form of both high-definition imagery and geospatial insight.


LBx: From an operational perspective, how much of an increase in speed are you getting?

JASON: Across the whole process we are seeing a 15x improvement in performance. Orthorecti cation is a barrier to entry for imagery analysis. For example, with traditional CPU technology, what would take about 15 minutes to orthorectify an image, with GPU technology can take one minute. When lives depend on it, those minutes really count.

The same technology that is used to draw high quality images to your screen (think video games, and movies) can be used to process and derive information from input imagery. Unlike video game developers, we don’t need to generate an image; that’s what the satellite is for. Instead we need to be able to apply algorithms to existing images and derive more intelligence out of the images.

FIGURE 2. This illustrates the pixel processing process and the time savings in orthorectification and pan-sharpening by employing GPU technology.
FIGURE 2. This illustrates the pixel processing process and the time savings in orthorectification and pan-sharpening by employing GPU technology.

LBx: How does this all t into the evolution of satellite imagery and geospatial applications?

JASON: Looking across the historical timeline of commercial satellite imagery, the initial emphasis was on accuracy… How accurate is the information? Accuracy is now no longer an issue. Next was how much information can you collect. DigitalGlobe now has the most complete imagery of all the commercial satellite imagery providers.

75% of all commercial imagery is collected by DigitalGlobe – that includes a two billion square kilometer archive, and a collection capacity of more than 700 million square kilometers per year.

Phase 3 in this evolution is currency – the refresh rate of the images… How new is the information? These 3 phases or levels of the business have all improved dramatically. The next phase of the industry’s evolution is extracting information from the imagery. In other words, we are entering the era of information and insight from satellite imagery and we believe that GPU computing is critical to this next phase.

KEVIN: Many in the geospatial field erroneously think that a GPU architecture requires complex programming. Our CUDA (Computer United Device Architecture) parallel programming environment provides all the tools necessary to help developers port applications to the GPU so that they can take advantage of GPU performance gains. Hundreds of thousands of application developers use CUDA, and more than 500 universities and training centers around the world teach CUDA programming, making it the world’s most popular solution for programming GPUs.

GPU computing is growing in popularity in many elds, and will lead to new innovations on geospatial applications. We’ve only just scratched the surface of what is possible.

FIGURE 3. This is a low-resolution multi-spectral (e.g. color) image. Each pixel represents 2 meters.
FIGURE 3. This is a low-resolution multi-spectral (e.g. color) image. Each pixel represents 2 meters.
FIGURE 4. This is the high-resolution panchromatic image (e.g grayscale). Each pixel represent .5 meters
FIGURE 4. This is the high-resolution panchromatic image (e.g grayscale). Each pixel represent .5 meters

 

FIGURE 5. This is the resultant image from the pan-sharpening process. Color information from the low-res multi-spec- tral image is merged with the high-resolution panchromatic image to create a high-resolution color image. Graphics courtesy of DigitalGlobe.
FIGURE 5. This is the resultant image from the pan-sharpening process. Color information from the low-res multi-spectral image is merged with the high-resolution panchromatic image to create a high-resolution color image. Graphics courtesy of DigitalGlobe.