|
{intl-found}: http://www.headwave.com/article/articleprint/65/-1/19/
Or navigate from www.headwave.com 
News and Events
Newsroom
|
SEG 2007 Special Coverage
Supercharging Seismic Processing with GPUs
by Michael Feldman
HPCwire Editor
http://www.hpcwire.com/hpc/1803161.html
In the oil and gas industry, seismic data processing has become an essential tool in the energy production workflow. By employing the latest software and hardware, raw seismic data can be used to create revealing images of the geological structure of the earth's subsurface. These can then be used by geoscientists to find the most likely places to drill for oil and gas. Because of the enormous volumes of data generated by seismic exploration, the speed and accuracy of transforming the data into images is directly related to the amount of compute power that is available.
At the heart of these calculations is vector processing, which attracted early developers of seismic data processing to seek out vector-based machines. In the 1970s, one of the biggest users of supercomputers, like the IBM 3838 array processor and the original Cray vector machines, was the oil and gas sector. Following the rise of high performance computing clusters, the industry rewrote all their codes to run on standard commodity hardware.
But as the demand for real-time seismic analysis grows and as both power and space become issues in the datacenter, the oil and gas industry is looking at more compute-dense, energy-efficient solutions. Especially within the past couple of years, there has been growing interest in coprocessor acceleration, which uses FPGAs, GPUs, Cell processors, and ClearSpeed boards to perform the most compute-intensive parts of seismic processing.
One company that is offering solutions along these lines is Houston-based Headwave Inc., a company that sells software tools for seismic analysis. They are one of the first vendors to use GPGPU (general-purpose computation on GPUs) for commercial purposes. Headwave is employing NVIDIA GPUs to filter and visualize multi-terabytes of raw seismic trace (prestack) data. The latest graphics devices offer as much as 500 gigaflops of processing power.
According to Steve Briggs, Headwave's VP of Integration and Deployment, using the vector processing capabilities of GPUs enables interactive prestack investigation and analysis. Briggs, who came to Headwave in mid-May, was an early evangelist of GPGPU while at HP, his former employer. While there, Briggs saw the potential of GPUs for augmenting a wide variety of technical computing codes.
"One of the things that's exciting about GPUs is that you get vastly more compute power for the same wattage and density," says Briggs. "They're running out ahead of what the standard CPUs are achieving."
By delivering 3D visualization of prestack data, Headwave has managed to automate away some of the manual labor of examining the raw data with human eyeballs. The company has developed algorithms that compress the prestack information (30-50 times the amount in the poststack), essentially transforming multiple terabytes of data into something that is visually useful. Areas as large as 20x20 kilometers, and looking down to a depth of 20,000 feet, can be accommodated. Briggs says being able to visualize this information interactively enables oil and gas companies to locate the "interesting" spaces that much faster.
Energy firms spend on the order of $4 to $7 billion per year on data acquisition and analysis. Companies like WesternGeco, Fugro MCS, and Petroleum Geo-Services (PGS) all gather and process data on their own, and sell it to multiple oil and gas companies. Lately these companies have relied on what are essentially seismic processing farms, consisting of tens of thousands of CPUs, to continuously process data. As the price of oil and gas rises, companies are willing to explore more marginal areas and look for reservoirs at greater depths to find deposits. This means even greater volumes of data must be collected and sifted through.
As a result, no one can ever afford all the CPUs that they would like. That's where the GPU comes in. Headwave has been able to use NVIDIA hardware, hooked up to standard PCs, to accelerate application performance by a factor of 100 compared to CPU-based systems. Headwave software supports the current NVIDIA FX 5600 and GeForce 8800 family of chips and expects to support subsequent versions of NVIDIA's Tesla line of GPU hardware as it's developed. NVIDIA developed Tesla specifically for the high performance computing market. The new devices are designed to connect to both workstations and standard servers.
"Customers have talked about clustering them, because more is never enough," explains Briggs. "So there will be clustered GPUs in the future."
Today in oil and gas applications, much of the data gathering is done using up to 18 bits -- 14 bits of data and 4 bits of (amplifier) gain. Therefore, the current 32-bit implementations of GPUs are more than adequate for the job. However, for some workloads, more complex computations are required. Examples include applications that measure oil viscosities and sheer strength, or that model the oil/water/gas interfaces. In this realm, 64-bit precision becomes very important. NVIDIA plans to implement double-precision GPUs early in 2008; presumably AMD/ATI will not be far behind. Once this occurs, GPUs will represent a much more general-purpose accelerator solution -- not just for oil and gas applications, but for all scientific computing workloads.
The real problem with GPGPU today, or for that matter with any computational accelerator, is the difficulty of programming the devices. Headwave went out and hired a number of programmers from the gaming industry to deal with the graphics hardware. Briggs notes it was "easier to teach them about seismic processing than to train a seismic physicist how to program." He does admit that NVIDIA's CUDA C programming environment is making a difference, since it presents a standard API that doesn't change from one processor generation to another.
But what Briggs would really like to see are industry-wide APIs. He thinks GPGPU is destined to take off over the next couple of years, but standards would help move that process along. "You can be competitors, but you'd better figure out standards," says Briggs. "If you can figure out industry standards, that will grow the market."