PDF《Wayne Wolf》

Embedded Computer Architectures in the MPSoC Age Wayne Wolf Dept.

of Electrical Engineering Princeton University [email protected] 1. Introduction Embedded computers are no longer used as simple controllers. Instead, high-performance embedded processors perform complex algorithms and are linked together to form multiprocessors. Embedded computing provides students different take on computer system design because of the requirements imposed on these systems: ? Embedded computing systems generally require real-time performance. Real-time and average- time performance are very different animals. ? Battery-powered embedded systems must meet very stringent energy requirements [Aus04]. ? Although the software in embedded systems can be changed to optimize the overall system, the software must also meet the specifications of the application. As such, an architecturally-oriented embedded systems class emphasizes somewhat different concepts than a traditional, general-purpose computer architect- ure class. An embedded computing architecture class must use a methodology to help students quickly get their hands around an unfamiliar application. They must explore a broad range of architectures. They should also explore trade-offs between architectural modifications and software modifications to meet system goals. Distributed embedded systems, which are built from networks of embedded processors, are also widely deployed. This paper will concentrate, however, on systems-on-chips. 1. Multiprocessor Systems-on-Chips Multiprocessor systems-on-chips (MPSoCs) [Jer04] are, first of all, systems-on-chips. They implement complete applications on a single chip. (Although as Rich Page points out, most systems-on- chips are marketing single-chip solutions---they use one chip plus all the other chips that you need to make the SoC work.) MPSoCs are systems-on-chips that include one or more programmable processors. Systems-on-chips are generally adapted to the application to meet performance, power, and cost goals. Although modern VLSI fabrication technology provides us with very large chips, applications keep getting larger. Some markets are large enough that specialized architectures are inevitable and desirable. Multiprocessor systems-on-chips try to balance specialization and programmability. Programmable processors allow the SoC to be programmed after fabrication;

MPSoCs are often referred to as platforms because they allow for many implementations of a given type of system. Programmability offers many advantages: the same chip can be used in several products, reducing product cost;

design tasks can be compartmentalized;

and the platform chip may have a longer shelf life than a highly specialized SoC. Because these are systems-on-chips, they generally aren'

t traditional symmetric multiprocessors. They may use hardwired function units in addition to programmable processors. They may use several different instruction sets. They may have non-uniform memory spaces supported by asymmetric networks. Many multiprocessor systems-on-chips are now available for several types of applications: ? Mobile multimedia requires both high performance and low energy consumption. The ST Nomadik and TI OMAP architectures are MPSoCs that provide specialized architec- tures for audio, video, and communications. ? Home multimedia is not as tightly constrained on power as mobile multimedia but requires very high performance for applications like HDTV. The Philips Nexperia architecture is a well-known MPSoC for set-top box applica- tions. ? Networking requires very high performance and provides some opportunities for specialized parallelism. Network processors from Intel, Cisco, and others use hetero- geneous architectures to process packets at high rates. 3. Architectural Challenges Embedded computing and MPSoCs make for a full employment act for computer architects. We are in no danger of running out of applications that can make use of large amounts of computing power and that can support the design effort required to create an efficient application-specific platform. Several specific challenges flow out of our continuing need to design MPSoCs. Configurable processors, such as those provided by Tensilica, allow the SoC designer a convenient way of quickly building processors with customized instruction sets. One area in which designers need help is figuring out which instruction set extensions should actually be implemented. Another important goal is figuring out how to connecting configurable processors into multiprocessor networks. Hardware/software co-design [DeM01] is another way to increase system performance for a particular application. Accelerators, when properly designed, can significantly and efficiently increase performance. However, the application must be carefully analyzed to be sure that an accelerator actually improves overall performance. Heterogeneous multiprocessors for embedded applications generally implement pipelines of processes. Our own smart camera system [Oze05] is an example of a pipelineable application. The smart camera processes video in real time, using a number of distinct steps. The amount of work performed by these stages is generally data dependent and buffers are required to smooth out rates. As video data is processed, it is boiled down in size so that data ........