Pushing the GPU Envelope

By Doug Barney

French-based GPU-Tech recently announced Ecolib, a new set of libraries allowing developers to tap into the power of today’s sophisticated Graphical Processing Units. In fact, the one-year old company already has partnerships with ATI, NVIDIA, and AMD to help ensure compatibility.

According to GPU-Tech, more can be gained by directly exploiting, or writing to GPUs than exploiting the CPUs themselves, such as through clustering.

The idea is not entirely original to GPU-Tech. In fact, Peakstream, profiled a month ago in this newsletter (click here for the story) uses a similar technique, Stream programming.

Instead of writing so that operations happen serially, stream programming assumes that there are multiple processors, cores, or clusters, such that application processing is done in parallel.

GPU-Tech has several ways for programmers to go. They can write in C++ directly to the GPU-Tech API. This offers the greatest performance boost, but also takes the most time. Porting your app through the Ecolib libraries offers a faster approach. According to the company, “these libraries regroup algorithms used in numerous sectors like finance, oil and gas, aeronautics, life sciences, meteorology, acoustics, etc.”

Libraries include signal and image processing, basic linear algebra, computing of options and VaR, generation of random numbers, and resolution of linear system equations.

Finally GPU-Tech itself offers custom porting and programming services. The company’s approach is useful “for all algorithms using floating points in simple of double precision or using whole numbers.”