A new access to aerodynamic number-crunching:

Graphics card outperforms supercomputer

Computed turbulences behind a ball

28.11.2008, Press releases

Although computerized simulations save engineers a lot of time, they also bring their own challenges. Conventional computers are unable to deal with complex calculations and supercomputers are rare and costly. A student research project carried out at the Technische Universitaet Muenchen’s (TUM) Institute of Aerodynamics could help solve this problem. The project’s revolutionary findings reveal how conventional graphics cards can clearly outperform expensive supercomputers when performing complex calculations.

Vortex flows develop in the wake of all vehicles, absorbing energy and generating noise and vibrations as they go. Investigating these flows is just one area of research at the TUM’s Institute of Aerodynamics. An important research tool is the so called computational fluid dynamics (CFD). “CFD runs numerical simulations on irregular flow streams. It is an important supplement to wind tunnel experiments, particularly in the case of physically complex flows”, explains Professor Nikolaus A. Adams, Full Professor and Chair of Aerodynamics at the TUM.
“A typical simulation involving an extremely simplified vehicle model comprises 48 million three-dimensional volume elements, and requires over 102,000 time steps. A supercomputer costing several hundred thousand euros needs almost 60 hours to calculate this kind of simulation in full,” details Dr. Thomas Indinger, head of the Car Aerodynamics Group at the chair.
Yet the same task can be performed much quicker on a system that only costs between one and two thousand euros. And the secret behind this cost-effective, high-speed simulation solution is nothing other than a conventional graphics card. Thanks to their massively parallel architecture, graphics processing units (GPUs) can complete computationally intensive tasks considerably faster than traditional central processing units (CPUs).
Eugen Riegel, a 4th year student in aerospace technology, hit upon this idea after reading an article in the German computer magazine c’t. The piece in question described how GPUs were being used for science and research in conjunction with CUDA, a programming language developed for graphics cards. “I then decided to perform simulation computations using a NVIDIA GeForce 8800 GT graphics card with 512 Mbyte of memory and make these tests the subject of my term paper”, reports Riegel. The results were astounding. This mid-range graphics card, available from as little as 100 euros, computed 7 times faster than conventional CPUs.
In contrast to CPUs, graphics processing units could not previously be programmed. Now, the advent of GPU programmability is positioning graphics processors as high-performance computing systems. This move was made possible by the company NVIDIA, which developed CUDA (Compute Unified Device Architecture), a C/C++ based programming language for graphics processors. CUDA is freely available and the software can be downloaded free of charge from the company’s website. Graphics cards draw their considerable computational power from the parallel data processing architecture on the graphics chip, which provides significantly more transistors for calculations than conventional CPUs.
Reigel’s project supervisor Dr. Thomas Indinger also recognizes the great potential of graphics processors for scientific and research purposes. “This project has shown that thanks to their massively parallel architecture, graphics processors are able to perform compute-intensive tasks significantly faster than conventional CPUs. We see huge growth potential for GPU solutions, particularly in data- and compute-intensive basic research."
The TU Munich and NVIDIA have now entered into a partnership, whereby NVIDIA provides the Institute of Aerodynamics with graphics processors from its high-performance computing product line Tesla. Designed for continuous use in professional environments, the processors are equipped with up to 4 GB of memory and achieve speeds of up to 1 TeraFLOP. Scientists at the TU Munich will soon be using a Tesla system to perform flow simulations and are aiming for a 40-fold reduction in processing time.

Contact:

Kontakt: presse@tum.de

More Information

http://www.nvidia.de/object/tu_munich_de.html