Exploiting Detachability: A Non-Silicon Approach to Polymorphism

General purpose processing elements are designed to attain high overall performance across a wide range of applications with different workload characteristics. This greatly compromises the performance of the execution of individual applications. Moreover, with the commercial advent of chip-multiprocessors, the designspace of processing elements has greatly increased, rendering the performance of fixed architectures in the execution of arbitrary workloads more suboptimal than ever. The gist of this paper is the argument that the performance enhancement attainable through employing processing elements that are custom-designed for different workloads is greater than that ever attainable through conventional forms of implementing architectural polymorphism. Thus the full potential of customization will only be attainable if it were made possible to physically remove the processing element actively employed in the computer system and replace it with one that is customized to the workload characteristics of the application at hand. Through exploitation of the physical detachability of processing elements the architectural diversity of available configurations can scale unboundedly without degrading chip yield or the performance of individual configurations. These are critical features that are absent in previously proposed approaches to customization. We qualitatively argue that through this approach, a significant form of performance enhancement is achievable that will otherwise remain untapped.