Development and validation of Nessie: a multi-criteria performance estimation tool for SoC

The paper presents Nessie, an original multicriteria exploration tool able to decide on functional and nonfunctional aspects of SoC's based on the application and platform descriptions and the costs (energy, execution time, area, ...). Indeed, the different performance prediction tools we can find in the literature, aiming at helping designers to take decisions and find efficient solutions, have some limitations that we wanted to overcome. To validate our tool, we have chosen real case study. In particular, we present here an example based on a 3MF MPSoC platform executing a real time AVC/H.264 video decoding extended to 3D-stacked architectures. Results given by Nessie, based on the average wire power consumption, have reinforced us in the utility of such a tool to help designers to take decision, especially in huge design spaces like embedded SoC ones.

I. Introduction

In the nowadays so constrained embedded electronics world, the increasing complexity of the applications and the platforms (heterogeneity, IP and communication infrastructure possibilities) makes rapid and optimal decisions even more complicated for designer to satisfy the specification and the time-to-market constraints. In typical industrial top-down design flows, automatic tools appear generally at low abstraction levels and often explore one sole aspect of the system (functionality or architecture) taking into account one type of constraint (cycles, energy). Decisions made at higher abstraction levels are more devoted to experts that make intuitive assumptions about the design and the final performances and restrict the exploration space to be able to make quickly first decisions, sometimes without information about the counterpart of the system (functionality or platform). This leads inevitably to non-optimal solutions and long design time. To predict performances from the early stages of the flow, several tools have been developed proposing some solutions to guide the designers. The state of the art in the domain, presented in the next part, has driven us to elaborate our own original multi-criteria high level exploration tool, Nessie, more flexible and general. The validation of this original tool constitutes the second phase of the work and consists in simulating a real case to observe the feasibility to explore a given design space and generate relevant solutions for the users in few times. In this paper, we thus present a case study based on a real time video decoding AVC/H.264 application running on a 3MF MPSoC. While three functionality mapping scenarios with three data resolutions were considered, we added an extra degree of freedom by placing the hardware components on 3D stacked architectures. We have estimated for different functionality/platform possibilities the average power consumption gain we can achieve, in particular on the Network-on-Chip, the communication infrastructure. The obtained results have shown the interest of such a tool.