Superprocessors and Supercomputers

Abstract

In this article, we describe current state-of-the art processor designs, the design challenges faced by technology, and design scaling slow-down, problems with the new design paradigms and potential solutions as well as longer-term trends and requirements for future processors and systems. With technology and design scaling slowing down, the processor industry rapidly moved from high-frequency designs to multi-core chips in order to keep delivering the traditionally expected performance improvements. However, this rapid paradigm change created a whole new set of problems for the efficient usage of these multi-core designs in large-scale systems. Systems need to satisfy an increasing demand in throughput computing while at the same time still growing single-thread performance significantly. The increase in processor cores poses severe challenges to operating system and application development in order to exploit the available parallelism. It also requires new programming models (e.g. OpenCL*). Furthermore, commercial server systems are more and more enriched with special-purpose processors because these specialty engines are able to deliver more performance within the same power envelope than general-purpose microprocessors for certain applications. We are convinced that future processors and systems need to be designed with tight collaboration between the hardware and software community to ensure the best possible exploitation of physical resources. In the post-exponential growth era, hardware designs need to heavily invest in programmability features in addition to the traditional performance improvements.