By Allan Maurer
ATLANTA—Scientists, engineers, and financial analysts often use software called MATLAB to handle their massive data sets. But switching their data to run on one of the newer graphics processing units and using the AccelerEyes GPU toolbox can run their applications 50 to 100 times faster, the young company says.
Founded in June 2007, AccelerEyes won the 2008 Georgia Tech Business Plan competition and $50,000 in prizes over 29 other companies.
CEO John Melonakos tells TechJournal South the company is self-funded by its founders who came from its parent firm, DivEyes. The firm has four fulltime and six part time employees and is hiring programmers with specific skills, he says.
“We started AccelerEyes because we realized there is a big niche for visual computing.” When analyzing data such as DNA/RNA unfolding, a very complex process, MathWorks’ MATLAB users do not want to see data resolved as a numerical answer, but rather, want to observe the process visually,” says Melonakos. The same is true of other engineering and scientific disciplines.
MATLAB code is used by the aerospace and defense, automotive, biotech, pharma, and medical device, communications, electronics, financial services, industrial automation and semiconductor industries.
Also, Melonakos points out, that powerful graphics processing units (GPUs) driven by the video game market have or emerged or are emerging.
These have the potential to bring parallel computing to the PC, competing with less powerful CPUs.
Intel, for instance, has 1200 people working on its new Larrabee GPU expected to come out in 2009 or 2010. AMD is working on a GPU called Fusion. And Nvidia has developed a new GPU for speeding up complex computational code such as that used by MATLAB customers.
“All three of these companies are aware of this huge problem that’s existed in the marketplace with multi-threading and they’re all trying to overcome it,” says Melonakos.
Currently, he notes, most applications cannot even take advantage of the dual and quad core central processing units (CPUs) in PCs. These powerful GPUs, on the other hand, offer much broader parallel computing potentials of 256 or 1,000 cores.
The main problem—one AccelerEyes solves for MATLAB users—is that the programming necessary for applications to take advantage of multi-threaded parallel computing is “very, very difficult,” explains Melonakos.
“To be able to use them, a programmer has to tell the application how to use, or design into the application the ability to use those cores. The program constructs to do that are very difficult and the average programmer is unable to do it.
“What we’re doing is trying to overcome the problem for scientists and engineers,” he says. “We enable MATLAB code to run on these emerging GPU chips.”
Even though the programming task of taking advantage of 256 or 1,000 cores is even more daunting than dealing with dual and quad cores, the company’s Jacket product hides all that complexity from the MATLAB end user.
By simply casting MATLAB data to a GPU buffer via its ‘gsingle’ command, the user can simply declare that the code should run on the GPU. The rest of the functionality is hidden from the user.
“Jacket enables MATLAB code to run on the PC as is, so they don’t have to do any weird multi-threading programming so their applications will run on these highly parallel chips. You can imagine the speedup.”
One benefit is that users get much higher speed visualizations, which is important for untangling genetic code data and other complex computational data sets.
AccelerEyes already has thousands of Beta users and plans to start selling the licensed software at the beginning of next year. Depending on how sales go, says Melonakos, the company may seek additional funding to accelerate its growth.
On the Web: www.accelereyes.com
© 2008, TechJournal South. All rights reserved.
