The number of available commercial platforms is constantly increasing. The choice of an architecture that fit as much as possible the requirements is therefore more and more complex. This is even more real with the availability of recent multiprocessors architectures. As a consequence, methodologies with their associated tools are required in order to quickly evaluate future platforms, so that choices can be made early in the design flow. So far, evaluating either the performance or the power consumption of a dedicated platform was performed through executing benchmarks and applications on this platform. In this thesis, a new methodology with its associated tools, called FORECAST, is proposed to model both the hardware and software of a system, and then to estimate its performance and its power consumption. Our methodology is based on efficient models, easy to characterize using only information provided by constructor datasheets. Moreover, our approach is able to automatically generate an executable code of the system that can be simulated on the host machine. This simulation allows a rapid execution of multiple test cases. Our approach is therefore well adapted for performing architecture exploration. A lot of experimentations have been performed using our tool FORECAST for different applications (H.264 video decoder, radio application, benchmarks. . .) and different hardware platforms. Results obtained both in performance and in power consumption have then been compared with existing platforms (OMAP3, OMAP4, i.MX6, QorIQ. . .), but also with two collaborative projects, OpenPeple (ANR) and COMCAS (Catrene), dealing also with performance and power estimations. The comparison demonstrates the accuracy of our approach as the estimation is always below a 20% error margin. These experimentations have also shown that our methodology provides a very efficient ratio between the modeling effort and the accuracy of the estimations.
