Programmable embedded systems are ubiquitous nowadays, and their number will 
further increase with the emergence of Ambient Intelligence. One of the 
first challenges for embedded systems is mastering the increasing complexity of 
future Systems on Chip (SoC). The complexity will increase relentlessly because 
the applications will become more and more demanding with the algorithmic 
complexity growing exponentially over time. Performances can only be reached 
by using all forms of parallelism. This will bring the challenges of 
designing multi-core systems using all possible levels of parallelism 
to reach the required performance density, of extracting all the 
parallelism from the application(s) and of efficiently mapping this
efficiency to the hardware.

But for most embedded systems, a main challenge is in having sustained 
performances, not peak: guaranteed performances, and predictable timing 
behavior are important, together with Quality of Service, safety, reliability 
and dependability. The notion of time is key in embedded systems, and most of 
the current methods and tools, inherited from mainstream computer science, 
did not really cope with this extra requirement.

The new technology nodes (65nm, soon 45nm) will also bring their own 
challenges: the global interconnect delay does not scale with logic, the 
leakage power will be more and more important, and the increasing variability 
of components will be major problems for the design of complex systems. 
Assembling systems with {\em unpredictable} elements will increase the 
global system unpredictability.  Also, systems are not really designed with 
{\em separation of concern}in mind, and due to shared resources, a slight 
change can have a drastic impact.

Major breakthroughs will be required in compiler technology and in mapping 
tools, both in term of correctness and in terms of performance to achieve an 
efficient use of resources (performance- and power-wise), but also for the 
debug, validation and test of the system. Systems can no longer be verified 
with simulations, and we will need new validation approaches.  Otherwise, the 
unpredictability and unreliability due to the combination of use cases will 
make the systems practically unusable.