We’re all well aware that the raw computing power of the modern graphics card is well beyond that of most CPUs. The parallelized structure of the GPU is very well suited for highly repetitive and number-crunching tasks, while it on the other hand may not be as flexible as a CPU is. Researchers in Belgium, more precise the research group ASTRA, part of the Vision Lab of the University of Antwerp, decided to investigate if it was possible to build a desktop supercomputer by using the power of multiple GPUs. The reason is because the ASTRA group mainly focuses on the development of new computational methods for tomography. Tomography is used to create 3D X-ray pictures.
Computing these 3D pictures take a very long time using a regular PC, days to weeks the least, when working with high resolution 3D pictures (E.g. 1024 x 1024 x 1024 px). You can always connect multiple PCs and spread the workload across, but that would become rather expensive rather fast. So instead they decided to develop FASTRA. A desktop supercomputer that takes advantage of the fact that we have hardware specializing in rendering pictures and 3D images; the so called graphics card.
They first decided that they wanted to use NVIDIA’s latest dual-GPU card GeForce 9800GX2, because of CUDA, and then sought out to find a motherboard with four PCIe x16 slots with dual spacing. They found only one board that qualified; MSI K9A2 Platinum. It didn’t matter that much that it was an AMD motherboard since the CPU wouldn’t be doing much anyway other than making sure the graphics cards were happy.
The remaining hardware consisted of an AMD Phenom 9850 Black Edition, 8GB (4x2GB) Corsair TwinX DDR2 PC-6400, Samsung Spinpoint F1 750GB, Scythe Infinity for cooling, Thermaltake Toughpower 1500W Modular power supply and of course the four GeForce 9800GX2 from MSI. All of this was packed inside of a Lian-Li PC-P80 Armorsuit case.
Since CUDA doesn’t support Vista yet, and they wanted 64-bit support, they decided to go with Windows XP-64. They also feared that the driver support for Linux might become an issue, which ruled that Linux out of the picture. You might be wondering why they decided to go with a motherboard based on an AMD chipset, when it obviously lacks any support for SLI, and the reason is simple. They don’t need SLI. SLI is only for communicating in-between GPUs, while with their setup, all communication is GPU-CPU.
They sat down and compiled software that would distribute the workload across the GPUs and then decided to compare their new system to the local supercomputer, CalcUA. A 3.5 million euro supercomputer built back in March 2005, sporting 512 dual-core Opterons. The results are pretty amazing. A desktop PC costing less than 4,000 euro was able to deliver performance near the €3.5M supercomputer for a forward project and even beyond it when doing a complete reconstruction, which is what they are really interested in. They also discovered that they could easily overclock the cards by 20% and gain a stubstantial amount of additional computational power;
