Real-time Data Processing

Printer-friendly version

Development of a HPC system for a Fraunhofer Institut

Development of a HPC system for real-time processing
Based on a concept of the Fraunhofer FKIE (Fraunhofer Institute for Communication, Information Processing and Ergonomics), FluiDyna developed a real-time system for broadband processing of measurement data for this Institute in early 2010, applying NVIDIA’s computing architecture "CUDA."

This system allows a continuous processing of ten analogous channels which are digitized simultaneously with 70 MHz with no loss of information and subsequently undergo a FFT (Fast Fourier Transform) on the graphic cards.

The challenge
Prior to the introduction of the new system, the data packages to be processed had been distributed among the ten computers of a standard CPU computer cluster on which FFT processing had been conducted offline, thus time-delayed. This process had always taken a few days.

Therefore, the project was characterized by two specific challenges: First, in order to obtain the aspired to increased power density levels, the required computing power was to be added to as few and as compact computing components as possible. Second, the system was to fulfil the real-time processing requirement during each of the following three stages:
• Data gathering (digitalization of analogous data and input into the computer),
• Data processing (FFT processing), and
• Output of results using a multi-cast method (data output into a high-performance cluster to be processed further at various work stations).

The Solution
In addition to very efficient A/D-transformer cards (whose FPGA firmware was also developed by FluiDyna), the core of the individually programmed supercomputing cluster was made up of two NVIDIA S1070 units comprising a total of eight GPUs.

On this supercomputing cluster, ten analogous channels can be scanned with 70 MHz with no loss of data and simultaneously undergo a FFT. For this setup, FluiDyna GmbH developed customized software which uses NVIDIA’s CUDA technology.

The Impact
Previously, a standard CPU computing cluster comprised ten computers and had needed several days to process data packages.

The new system enables the same amount of data packages to be processed using only two computers - each of which is equipped with four graphic cards - in a supreme manner in real-time, with a delay of less than one second. Due to its enormous efficiency and performance, the new super-computer cluster opens up a number of new potentialities. In addition to the significant power enhancement and time savings, there is also the major benefit of cost savings stemming from the much lower energy consumption when measured against a conventional CPU solution. Compared with a CPU system with similar output levels, total costs are now reduced to only one tenth.