'Cell University Challenge' winners selected
Wednesday 26th September 2007, 03:03:00 PM, written by Carl BenderTaking first prize ($10,000) in region 1 was a project by students Jayram Moorkanikara Nageswaran and Jeff Furlong from the University of California, Irvine, and Ashok Chandrashekar and Andrew Felch from the Neukom Institute for Computational Science at Dartmouth. The team successfully ported a vision algorithm derived from brain research to a cluster of three Playstation 3s, tasking the system with successfully identifying individual objects from within complex images. By the conclusion of the project, work that had taken their baseline 2 GHz Core2 Duo PC three minutes to perform was being completed by the Cell cluster in as little as one second, with an average speedup of 140x.
The algorithm functions by breaking distinct objects within a scene
down into line triplets; forms are then compared against alternate
images to determine a match. As the work done by the processor in this
task is similar in nature to what is required in language processing,
the hope is that progress in this vein will assist in leading to a
practical robot design capable of both understanding verbal commands
and self-navigating through unfamiliar terrain.
Other prize winners for region 1 were as follows:
Second place went to Marc de Kruijf of the University of Wisconsin, Madison for his work on MapReduce for the Cell BE. Proposed by Google as a parallel programming model for distributed large-scale data processing, Marc's implementation saw a 2.5x improvement over a 2.4GHz Core2 processor on computationally intensive applications. Moreover, performance was shown to scale linearly with the addition of SPEs, with runtime overhead a low 4 percent.
Third place was won by Jusub Kim, a student at the University of Maryland at College Park. Jusub created a volume ray-casting algorithm aimed at displaying imagery from MRI, CT, and 3-D laser scans in a desktop environment in realtime. Taking data points 256x256x256 in size, the algorithm was able to display a 256x256 ray-cast image at a rate of 15 frames per second on a single Cell processor. In contrast, the algorithm running on a 3GHz Intel Xeon was 100 times slower.
Fourth place went to Robert Hiramatsu and Jussara Kofuji at the University of São Paulo. In a re-implementation of OpenCV rapid object detection, a stump-based classifier algorithm was used to reduce data structure for classifiers. The work is relevant to applications such as facial recognition.
Although the winning projects for region 2 have also been determined, IBM is waiting to publicly recognize these participants at a later regional event. All four winners for region 2 hailed from Chinese universities, with the winning projects ranging from CT reconstruction to H.264 encoding.
Tagging
Related News
RWT explores Haswell's eDRAM for graphics
ATI shoots a Bolt through its GPU compute stack
AMD releases CodeXL 1.0
Travelling in Style: Beyond3D's C++ AMP contest
Analysis of Ivy Bridge Graphics Architecture at RWT
RWT analyzes Kepler's architecture
Nvidia 680 GTX (Kepler) Released
Microsoft Releases C++ AMP Open Specification
Nvidia's 2x Guaranteed Program