Invited Speakers
1 - George A. Constantinides, Imperial College, London, UK
"Efficient and Automated Fixed-Point Implementation"

This talk will provide insight into the problems and solutions related to automating the mapping of an essentially infinite precision algorithm into a fixed-point implementation. Such a mapping is often essential in order to achieve fast, low area, and low power implementations, particularly on FPGAs, where the overhead of floating-point is significant. Few commercial synthesis flows so far integrate such a design step, which is typically left to the designer to explore by hand, resulting in inefficient implementations.

We shall examine both the theoretical and practical issues surrounding this mapping. In particular, we will focus on two issues of practical importance: how to avoid overflow in a fixed-point implementation, and how to guarantee, or estimate, the arithmetic accuracy of a given fixed-point implementation.

It will then be demonstrated that FPGAs, and customised hardware in general, allow one to shape the precision of the datapath to the application of interest, avoiding the “one-size-fits-all” approach of the microprocessor, and opening avenues for significant area and power consumption reduction.

Finally, we shall examine some possible optimization methods to find the smallest area implementation satisfying guarantees on arithmetic accuracy, and provide some comments on the computational complexity of this problem.

2 - Stamatis Vassiliadis, Delft University of Technology, The Netherlands
"PISC: Polymorphic Instruction Set Computers"