Using Amdahl's Law for Performance Analysis of Many-Core SoC Architectures Based on Functionally Asymmetric Processors

摘要

Amdahl's law is a fundamental tool for understanding the evolution of performance as a function of parallelism. Following a recent trend on the timing and power analysis of general purpose many-core chips using this law, we carry out an analysis aiming at many-core SoCs integrating processors sharing the same core instruction set but each potentially having additional extensions. For SoCs targeting well denned classes of applications, higher performances can be achieved by adding application specific extensions either through the addition of instructions in the core instruction set or through coprocessors leading to architectures with functionally asymmetric processors. This kind of architectures is becoming technically viable and advocated by several groups, but the theoretical study of their properties is yet to be performed: this is precisely our goal in this paper. We use Amdahl's law to prove the performance advantage of using extensions for many-core SoCs and shows that the many-core architecture based on functionally asymmetric processors can achieve the same performance as the symmetric one but at a lower cost.