Acconovis is a young spin-off company from the Institute of Computer Science V at the University of Mannheim. It was founded in 2001. Acconovis hosts experts in massively parallel computer systems and specialises in applications such as scientific computing and industrial image processing for research and industry. Its core competency is situated in the development of high-performance HW/SW systems based on FPGA technology. Currently, the primary focus is put on intelligent high speed bioinformatics solutions.

Holger Singpiel explained to the audience that the Acconovis technology is based on interdisciplinary know how in the fields of computer science, mathematics and science, and new data-processing concepts with regard to application knowledge, high performance computing and parallel processing. The company's expertise in FPGA technology is unique since a number of software development tools as well as an intellectual property module library have been built.

The speakers also shed a light on the history and backgroun of FPGA computing. In 1985, Xilinx introduced the first field programmable gate arrays device and in the early nineties of the former century, systems such as Splash in Maryland, USA; DecPeRLe in Paris; and Enable at the University of Mannheim saw the light.

The state-of-the-art technology evolved from single to FPGA and coprocessors which are PCI-based. FPGA processors are situated between ASIC and µP processors.

The big advantage of FPGA technology is that it combines the ease of software with the speed of hardware. FPGA processors have been programmed for design entry, simulation, synthesis and P&R. There have also been efforts in developing high level design languages such as HW/SW co-design. FPGA processors are used in hybrid architectures such as Virtex-II Pro from Xilins and the XBluearchitecture from IBM.

According to the speakers, FPGA processors function as a hardware accelerator and they perform extremely well. Mr. Singpiel gave an example in which N-body simulations were accelerated with FPGAs, a project executed by the University of Mannheim and Tokyo University. In practice, this came down to one pipelined

arithmetic unit and one summation per clock cycle. As a result, the arithmetic fitted in 49% of Virtex-II XC2V3000-4 FPGA, the design frequency was 65 MHz and the floating-point performance included 3.9 Gflops.

The applications for which FPGA processors are used are interfacing and controlling, real time pattern recognition and image processing, data security or cryptography, medical applications, scientific computing, knowledge management, and of course bioinformatics. New computing concepts based on FPGA technology have the benefit to reduce the execution time of running a programme by a factor 10 to >100 and to increase at the same time the quality of the results, according to the speakers.

A very interesting scientific application to which FPGA technology can contribute a lot when it comes to programming, is the emerging field of genomics. In biology, researchers are dealing with DNA that is replicated and transcribed to RNA, which in turn is translated into proteins that modify and have to be localised in order to detect their function. To that purpose, scientists are comparing genes to gain insight in their meaning and function. Holger Singpiel mentioned BLAST (basic local alignment search tool) as an ideal instrument to perform this type of research.

The reasons why genes are being compared intensively are the eager to stay up-to-date with the inferred functions of proteins; to compare the theoretical capabilities of organisms and the different versions of production strains. Researchers are thus aiming at defining new targets for strain development.

Mr. Singpiel told the audience that at Acconovis the NCBI BLAST solution has been implemented. The original NCBI source code has been used but Acconovis researchers rewrote the time-consuming parts in order to execute them on the FPGA Co-processor.

This BLASTn solution has been benchmarked in a test where yeast genes have been compared versus the Arabidopsis genome. The execution time on a state-of-the-art PC was six hours and one minute. On the Acconovis HW "PC + 1 FPGA Co-processor" it was only seven minutes, which means a speed-up of 52 minutes.

The primer design for micro arrays is computationally very intensive, according to Mr. Singpiel given the multiple optimisation criteria, the amount of possible primers, and the database analysis. At Acconovis, the engineers have installed a complete parallel and pipelined implementation with a complexity reduction from O(n2 + m2 + mn) to O(1). The results were a 1024 spot micro array, an SW execution of 1920 seconds and a HW execution of 3,4 seconds. The speed-up was > 500 with 28 GOp/s.

Mr. Singpiel concluded his talk by highlighting the performance of a protein structure prediction, an example of the "ab initio" method and an implementation of 1998 with the Monte Carlo genetic algorithm and the pair potential technique. The result was a speed-up of factor 40. This application allowed the researchers to perform homology modelling and to generate more complex energy models.