# Conventional graphics cards beat supercomputers

Complex flow simulations at the Chair of Aerodynamics at the Technical University of Munich generally consist of 48 million three-dimensional volume elements and require over 102,000 time steps. So far, these numerical flow simulations have been calculated by supercomputers. However, the computing time available for large computers is scarce and also very expensive.

At the Technical University of Munich, an experiment showed that commercially available graphics cards beat modern supercomputers by a long shot in complex simulation calculations. Future flow simulations and other calculations are therefore carried out by researchers and engineers on NVIDIA graphics cards.

A test setup shows that graphics cards can also perform the extensive calculations and that they are even significantly faster than a supercomputer. In the experiment, no special graphics card was used, but a commercially available gaming graphics card from NVIDIA.

Anyone can buy such MSI graphics cards from Mindfactory or other shops for computer accessories on the Internet. For the test setup, not entire batteries from graphics card associations were installed, but only an NVIDIA GeForce 8800 GT with 512 MB memory – a graphics card that can be had for around 100 euros.

## Flow simulations supplement wind tunnel tests

One research task at the Chair of Aerodynamics at the Technical University of Munich (TUM) is the investigation of the wake vortex on vehicles. These swallow energy, cause noise and vibrations. An essential tool of this research is numerical flow simulation – Computational Fluid Dynamics (CFD). “With CFD, fluid mechanics problems are numerically simulated, which makes CFD an important addition to wind tunnel tests, especially for physically complex flows,” explains Professor Nikolaus A. Adams.

“A typical simulation on a highly simplified vehicle model consists of 48 million three-dimensional volume elements and requires more than 102,000 time steps. A supercomputer costing several hundred thousand euros takes almost 60 hours to fully calculate such a simulation, ”explains TUM researcher Thomas Indinger.

The same task can also be completed much more quickly – on a system that only costs one to two thousand euros. The secret of this inexpensive temporary exchange: The simulations are carried out using conventional graphics cards. Due to their massive parallel architecture, graphics processors (GPU) can perform computation-intensive tasks many times faster than conventional main processors (central processing unit, CPU).

## Programmability of the graphics card:

The basis for the use of graphics cards for the calculation of complex simulations based on GPU is the free programmability of the graphics card – a feature that only CPUs had in the past. With the help of the programming language CUDA (Compute Unified Device Architecture) based on C / C ++, the researchers at the Technical University of Munich are able to adapt the graphics card to the desired requirements for the simulation calculations.

The significantly higher computing power of the graphics cards results from the parallelization of many data processing units on the graphics chip, which means that, compared to conventional CPUs, many more transistors are available for the calculation.

## Faster and cheaper:

The simulations of the so-called wake vortices on vehicles are extremely complex and important. These swallow energy and cause noise and vibrations. “A small, hundred-thousand-euro supercomputer takes around 60 hours to calculate such simulations, ” explains Dr.-Ing. Thomas Indinger, head of automotive aerodynamics. Compared to the conventional approach, the calculation using graphics cards is not only much cheaper, but also around seven times faster. The complete computer system that beat the supercomputer cost just one to two thousand euros.

## Great future for graphics cards in science:

Dr.-Ing. Thomas Indinger sees great potential in the use of graphics processors in science and research: “The experiment has shown that due to their massive parallel architecture, graphics processors can perform computation-intensive tasks many times faster than conventional main processors. We see great opportunities for the increasing spread of GPU solutions, especially in areas in which data and computation-intensive basic research is carried out. ”

The TU Munich and NVIDIA have decided to cooperate based on the experiment. NVIDIA provides the chair for aerodynamics with graphics processors from the high-performance computing product line Tesla, which is designed for continuous use in a professional environment and is significantly more powerful. At TU Munich, flow simulations will soon be carried out using a Tesla system. The goal of the scientists: to accelerate the calculations 40 times.

Simulations save a lot of time. But more complex calculations can quickly overwhelm conventional computers, and computing time on large computers is scarce and expensive. New results from a young Munich researcher could now help solve this problem: He used standard graphics cards for complex calculations. With the result that expensive supercomputers had to give up.

## Great potential

Indinger, who supervised the research work, sees great potential in the use of graphics processors in science and research: “The work has shown that, due to their massively parallel architecture, graphics processors can perform computation-intensive tasks many times faster than conventional main processors. We see great opportunities for the increasing spread of GPU solutions, especially in areas in which data and computation-intensive basic research is carried out. ”

The TU Munich and NVIDIA have now decided to cooperate. NVIDIA provides the chair for aerodynamics with graphics processors from the high-performance computing product line Tesla, which is designed for continuous use in a professional environment. The processors have up to four GB of memory and offer computing power of one teraflop. At TU Munich, flow simulations will soon be carried out using a Tesla system. The goal of the TUM scientists is to accelerate the calculations 40 times.

## Seven times as fast

Eugen Riegel, an aerospace student, came up with the idea. In a computer magazine he read a contribution on the use of graphics processors in science and research as well as the programming language CUDA developed for graphics cards. “I then made the simulation calculations with a graphics card NVIDIA GeForce 8800 GT with 512 Mbyte memory the topic of my semester thesis,” he reports. The result was astonishing: with the help of the mid-range graphics card, which is available from 100 euros, Riegel was able to accelerate the calculations by 7 times compared to the conventional procedure using the CPU.

The basis for using GPUs as a high-performance computing system is their free programmability – a feature that only main processors had in the past. The programming language CUDA (Compute Unified Device Architecture) based on C / C ++ was developed in order to be able to program graphics processors.

CUDA is freely accessible, the software is available for download free of charge from NVIDIA. The high computing power of the graphics cards results from the parallelization of many data processing units on the graphics chip, which means that a lot more transistors are available for the calculation compared to conventional CPUs.