The same inexpensive hardware that powers Grand Theft Auto, Call of Duty and other calculation-intensive video games is being adapted to augment multiprocessor, multicore computers. Selling for as little as $1,500 each, graphics accelerators are helping some calculation-heavy applications run significantly faster.
But even if your business doesn't model financial markets, sift DNA sequences or map the earth's crust in search of oil, accelerators may someday boost the speed of your applications, too.
As chipmakers reach the limits of power density, they can no longer rely on boosting clock speed to increase processor performance.
So, instead of attempting to drive one chip faster, chipmakers are focusing on widening the road. Adding more "cores" to each chip makes it possible to perform more calculations simultaneously.
To tap the full processing power of this wider road, programs need to execute multiple streams of work simultaneously: A complex problem programmers have struggled with for years.
For the typical business application today, achieving such rare "parallelism" isn't necessary—at least, not yet. However, some calculation-intense programs can get tremendous speed boosts through parallelism, today.
Even though eight-core systems are common and growing less expensive every day, clustering enough of them to get supercomputer speeds would result in a costly, complex system. The graphics accelerators that power video games, however, can have more than 100 cores on single board—and they're inexpensive, too.
"Video games involve computing highly-detailed, megapixel images 60 times a second, and parallelism is key to performance in this compute-intensive task," says Glenn Lupton, engineering team leader for HP's high-performance computing accelerators program, part of the Catalysts for HPC Innovation initiative.
"If you can offload the tremendous number of calculations it takes to simulate molecules bouncing around, for example, and make that application run two, or 10, or 20 times faster, that's an amazing thing," says Kent Koeninger, technology and product marketing lead at HP.
That speed boost can mean the difference between getting results in weeks and getting them in hours or minutes, says Ryan Schneider, chief technology officer for Acceleware
, a participant in HP's Catalysts for HPC Innovation.
However, retrofitting applications to run on accelerators has proven a daunting task. In the past, programmers had to learn (and adapt) highly-specialized, graphics programming languages to their needs. As each new generation of accelerators becomes more easily programmed, that task is easing.
Many accelerator cards now can be programmed using general purpose, C-like languages. And, Acceleware even allows many scientific, engineering and finance applications to run on accelerator-enhanced platforms without reprogramming.
Other specialized chipsets are also being pressed into service, including field-programmable gate arrays (FPGAs) and application-specific integrated circuits (ASICs), offering further possibilities.
Accelerators can offer speed boosts to some applications, but not without substantial investments in reprogramming. That's why HP's High-performance Computing Accelerators Program offers expert advice, testing and benchmarks on how to use accelerators, which applications can offer the best payoffs and what techniques are proving useful in this fast-emerging field.
Collaborating with customers on the leading edge of HPC acceleration also feeds back into HP's server design, making it easier to implement these technologies. For example, graphics cards can process billions of floating point operations (gigaflops) per second, but only if that data is delivered at the same high rates. That's why select HP ProLiant servers are available with super-fast, high-Wattage slots for plugging in high-speed, power-hungry accelerator boards.
"HP was the first to address this space and has worked to support and enable the technology to fullest extent on their platform," Schneider says.
While acceleration works with a small but growing list of applications, acceleration capabilities will eventually become part of every computer, technologists predict. As processor cores proliferate, even typical office applications will need to adopt some parallelism or performance will plateau (or in some cases, decrease). So, while parallelism and accelerators are on the cutting edge today, many more applications will benefit from them in future.
