Design Approaches to Application Specific Adaptive Processors

Ramesh Karri
Bell Labs/Lucent Technologies

Monday, March 23, 1998 
11am-12:15n
 Interschool Lab, 7th floor, Schapiro CEPSR Bldg.

Host: Steve Nowick

Abstract

Modern wireless and multimedia applications require high performance, low power, and inexpensive multiple functionalities. Conventional architectures fail to meet these requirements of algorithm flexibility and low complexity. While dedicated ASICs can be optimized to execute a given algorithm in real-time with very low complexity, they lack the flexibility to perform a range of algorithms. Software programmable processors, on the other hand, although highly flexible are complex and have low throughput. Consequently, rather than optimizing the architecture for a single algorithm or for all algorithms, we are investigating application specific adaptive architectures targeting classes of applications. Such architectures yield a suprior tradeoff between low complexity and high algorithm flexibility than either software-programmable processors or dedicated ASICs. In fact, almost all major processor manufacturers including Fujitsu (with their 86k line of programmable processors), Motorola (with their application specific programmable DSP processors) and LSI Logic are offering a comprehensive line of application specific adaptive processors that preserve all of the advantages of the special purpose ASICs while retaining the cost and flexibility afforded by the general purpose processors.

In this talk, I will present a coarse grain application specific adaptive processor architecture that has a low reconfiguration overhead, yields high reconfiguration rates and is adaptive in the presence of faults in the system. I will then discuss important issues and constraints that need to be considered while implementing adaptive processors for classes of applications. Finally, I will focus on computer aided design algorithms that automate the translation of high-level descriptions of algorithms into application specific adaptive processors.



Luis Gravano
gravano@cs.columbia.edu