A set of level 3 basic linear algebra subprograms
ACM Transactions on Mathematical Software (TOMS)
Instruction Set Extensions for MPEG-4 Video
Journal of VLSI Signal Processing Systems - Special issue on implementation of MPEG-4 multimedia codecs
Computer Architecture: Concepts and Evolution
Computer Architecture: Concepts and Evolution
Field-Programmable Custom Computing Machines - A Taxonomy -
FPL '02 Proceedings of the Reconfigurable Computing Is Going Mainstream, 12th International Conference on Field-Programmable Logic and Applications
The MOLEN Polymorphic Processor
IEEE Transactions on Computers
FCCM '04 Proceedings of the 12th Annual IEEE Symposium on Field-Programmable Custom Computing Machines
64-bit floating-point FPGA matrix multiplication
Proceedings of the 2005 ACM/SIGDA 13th international symposium on Field-programmable gate arrays
Validity of the single processor approach to achieving large scale computing capabilities
AFIPS '67 (Spring) Proceedings of the April 18-20, 1967, spring joint computer conference
Hi-index | 0.00 |
This paper reveals the evolution of the polymorphic architectures in the context of ever increasing computational demands of the user applications and the need for formal architectural abstraction covering the emerging reconfigurable technologies. The base architecture presented is the Molen polymorphic processor - a synergism between a general purpose processor (GPP) and a reconfigurable accelerator. Its operation is based on the co-processor architectural paradigm and employs the concept of the traditional microcode control. Experiments with popular media applications, such as MJPEG, MPEG-2, MPEG-4 suggest that a Molen architecture can provide speedups closely approaching the theoretically maximum obtainable figures, determined by Amdahl's law. However, while media applications are predominantly integer based, scientific applications, typically run on supercomputers, rely on floating-point arithmetic. Therefore, a natural evolution of the Molen platforms towards supercomputing, i.e., towards floating-point highly demanding operations, is in progress. The author presents some experimental results obtained for a typical supercomputing kernel, matrix multiplication, implemented as a Molen reconfigurable accelerator. These results demonstrate the advantages of the polymorphic approach agains traditional GPP based scientific computing in terms of high performance. Finally, the paper proposes author's vision for the future evolution of high-end polymorphic architectures.