Speaker: Florent Dupont de Dinechin - National Institute of Applied Sciences (INSA), Lyon, France

If an application does not compute accurately enough, somebody will complain. Conversely, as soon as an application computes accurately enough, chances are that it computes too accurately. Nobody will complain, but maybe they should? Computing too accurately is a waste of resources and power.
 
This talk is about computing just right: computing accurately enough, but not too accurately. More generally, it is about the opportunity, on FPGAs, to design the arithmetic operators for the application, instead of the other way around: in traditional programming we have to design the application around the few arithmetic operators offered by a microprocessor.
 
There is a qualitative aspect to application-specific arithmetic: this talk will review a few simple recipes that expose an infinity of useful application-specific arithmetic operators. Then there is a quantitative aspect, and we are back to "Computing just right": what accuracy is needed at each point of the computation? Answering this question leads to surprisingly difficult problems, for instance understanding what the code you just wrote was supposed to compute.