| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
|
Am 05.06.2013 08:53, schrieb John
Pryce:
This is certainly a nice idea.It seems to me it should be straightforward to implement CA (initially the binary64 version which IMO is always likely to be the most used one) as a commercially available co-processor chip -- similarly to how the first commercial version of 754 arithmetic was the 8087 co-processor. I look forward to someone taking up the challenge to create a CA chip. Prof Martin Berz has written (20091106) that CA would be very useful to his Taylor Model calculations, which have been crucial to various large projects such as designing the beam guidance for the Large Hadron Collider. Martin, how about your group taking on the CA chip? We had a similar idea twenty years ago. In 1993 we developed a chip for CA and the EDP as co-processor for the PC. See the picture on the poster. Very advanced design tools are available nowadays. So the implementation is indeed more or less straightforward. A general experience, however, speaks against doing it again. With design tools that are publicly available a processor only reaches about 10% of the speed a company would reach with its own optimized design tools. So the speed advantage that comes with the EDP (compared with a conventional computation of the dot product in floating-point arithmetic) gets lost by the less powerful design tools. In 2002 we developed another unit for the EDP in FPGA technology. Here the design is even simpler. It can fully be done at a mahtemetics institute. But the speed discrepancy to a comercial design is even larger. In the end the EDP must be incorporated into the ALU of the processor. On the side of the manufacturers I always met great openmindedness to improve their products. Of course, designing it causes some costs. It is on P1788 to build up enough pressure to get it done. "Shall" and "require" (as in motion 9) would be much better than "should" and "recommend". The hardware needed for the exact dot product is comparable to that for a fast multiplier by an adder tree, accepted years ago and now standard technology in every modern processor. The exact dot product brings a similar speedup for accumulations at comparable costs and it brings accuracy into computing. Of course I know that the standard IEEE P1788 should not require particular hardware. But I think for arranging the standard it is important to know that certain operations can easily be realized by simple and fast hardware. Best regards Ulrich -- Karlsruher Institut für Technologie (KIT) Institut für Angewandte und Numerische Mathematik D-76128 Karlsruhe, Germany Prof. Ulrich Kulisch Telefon: +49 721 608-42680 Fax: +49 721 608-46679 E-Mail: ulrich.kulisch@xxxxxxx www.kit.edu www.math.kit.edu/ianm2/~kulisch/ KIT - Universität des Landes Baden-Württemberg und nationales Großforschungszentrum in der Helmholtz-Gesellschaft |