Thèse de doctorat
Résumé : The work presented in this thesis targets the analysis and implementation of multi-criteria performance prediction methods for System-on-Chips (SoC).

These new SoC architectures offer the opportunity to integrate complete heterogeneous systems into a single chip and can be used to design battery powered handhelds, security critical systems, consumer electronics devices, etc. However, this variety in terms of application usually comes with a lot of different performance objectives like power consumption, yield, design cost, production cost, silicon area and many others. These performance requirements are often very difficult to meet together so that SoC design usually relies on making the right design choices and finding the best performance compromises.

In parallel with this architectural paradigm shift, new Very Deep Submicron (VDSM) silicon processes have more and more impact on the performances and deeply modify the way a VLSI system is designed even at the first stages of a design flow.

In such a context where many new technological and system related variables enter the game, early exploration of the impact of design choices becomes crucial to estimate the performance of the system to design and reduce its time-to-market.

In this context, this thesis presents:

- A study of state-of-the-art tools and methods used to estimate the performances of VLSI systems and an original classification based on several features and concepts that they use. Based on this comparison, we highlight their weaknesses and lacks to identify new opportunities in performance prediction.

- The definition of new concepts to enable the automatic exploration of large design spaces based on flexible performance criteria and degrees of freedom representing design choices.

- The implementation of a couple of two new tools of our own:

- Nessie, a tool enabling hierarchical representation of an application along with its platform and automatically performs the mapping and the estimation of their performance.

-Yeti, a C++ library enabling the defintion and value estimation of closed-formed expressions and table-based relations. It provides the user with input and model sensitivity analysis capability, simulation scripting, run-time building and automatic plotting of the results. Additionally, Yeti can work in standalone mode to provide the user with an independent framework for model estimation and analysis.

To demonstrate the use and interest of these tools, we provide in this thesis several case studies whose results are discussed and compared with the literature.

Using Yeti, we successfully reproduced the results of a model estimating multi-core computation power and extended them thanks to the representation flexibility of our tool.

We also built several models from the ground up to help the dimensioning of interconnect links and clock frequency optimization.

Thanks to Nessie, we were able to reproduce the NoC power consumption results of an H.264/AVC decoding application running on a multicore platform. These results were then extended to the case of a 3D die stacked architecture and the performance benefits are then discussed.

We end up by highlighting the advantages of our technique and discuss future opportunities for performance prediction tools to explore.