Published works confirm that parallelism at the data level is widely accepted as the most important performance leverage for the efficient execution of embedded media and telecom applications and has been exploited via a number of approaches the most efficient being vector/SIMD architectures. A further, complementary and substantial form of parallelism exists at the thread level but this has not been researched to the same extent in the context of embedded workloads. For the efficient execution of such applications, exploitation of both forms of parallelism is of paramount importance. This calls for a new architectural approach in the software-hardware interface as its rigidity, manifested in all desktop-based and the majority of embedded CPU's, directly affects the performance of vectorized, threaded codes.

The author advocates a holistic, mature approach where parallelism is extracted via automatic means while at the same time, the traditionally rigid hardware-software interface is optimized to match the temporal and spatial behaviour of the embedded workload. This ultimate goal calls for the precise study of these forms of parallelism for a number of applications executing on theoretical models such as instruction set simulators and parallel RAM machines as well as the development of highly parametric microarchitectural frameworks to encapsulate that functionality.