Antwerp (Belgium) - Recent advances in general purpose GPU computing are beginning to shift perceptions in supercomputing applications. Belgian researchers have assembled a relatively simple enthusiast PC with an emphasis on graphics processing capability, which beats a multi-million dollar supercomputer in its target application.

The desktop PC, called Fastra, was built with a focus on the development of new computational methods for tomography. Tomography is a technique used in medical scanners to create three-dimensional images of the internal organs of patients, based on a large number of x-ray photos that are acquired over a range of angles. As these 3D images can be quite large, advanced reconstruction techniques can sometimes require weeks of computation time on a regular PC. Which means that supercomputers are usually required to process computer tomography (CT) images.

The research group Vision Lab at the University of Antwerp came up with a different solution and constructed a PC that integrates four GeForce 9800 GX2 graphic cards (with a total of eight GPU cores) that runs CUDA-optimized tomography applications. The specifications include a MSI K9A2 Platinum motherboard, an AMD Phenom 9850 CPU, 4 x 2 GB Corsair TWINX DDR2 PC6400 memory, a Samsung Spinpoint F1 750 GB hard drive, a Thermaltake Toughpower 1500W Modular power supply unit as well as four MSI 9800GX2 cards. The researchers said that the cost of the system was less than 4000 Euro or about $5300.

It isn’t quite a tricked-out gaming system and the 3DMark06 score is just above what your average PC can manage to come with today (12,603 points). However, it is the CUDA application where this PC really shines. Compared to the 512-processor, $4.6M CalcUA supercomputer purchased in 2005, the PC can be more than a match: The projection of image slices took 23.4 seconds on the supercomputer and 35.1 seconds on the PC. The reconstruction of the slices was displayed after 67.4 seconds on the supercomputer systems and after just 52.2 seconds on the PC. The Vision Lab crew now believes that a real-time construction is possible through GPUs and is now building a cluster of such systems.

While it is an impressive example how GPUs can be applied in non-traditional ways, there are a few notes to be added. Of course, GPUs cannot replace traditional supercomputers, which still can be applied to applications with a broader range. Also, supercomputers usually carry huge memories, often in the Terabyte range, which cannot be matched by today’s GPU clusters. When we talk to scientists working with supercomputers and GPUs, they typically believe that future supercomputers will not completely transition to GPU clusters, but may develop into systems that consist of a traditional supercomputer structure as well as GPU capability.

An interesting side note about the Fastra PC is its motherboard. Eagle-eyed readers may have noticed that the MSI K9A2 Platinum board is not an Nvidia SLI-based board, but uses AMD Crossfire (780 chipset). The simple reason to choose this board may have been cost, but it is unlikely to impact the performance of the system: CUDA does not support SLI at this time, which means that the GPUs have to communicate with each other as well as with the CPU via PCI Express. The researchers claim that they have not seen any impact on performance and the GPUs apparently are scaling well.