%0 Journal Article
%T HoneyComb: An Application-Driven Online Adaptive Reconfigurable Hardware Architecture
%A Alexander Thomas
%A Michael Rščckauer
%A Jščrgen Becker
%J International Journal of Reconfigurable Computing
%D 2012
%I Hindawi Publishing Corporation
%R 10.1155/2012/832531
%X Since the introduction of the first reconfigurable devices in 1985 the field of reconfigurable computing developed a broad variety of architectures from fine-grained to coarse-grained types. However, the main disadvantages of the reconfigurable approaches, the costs in area, and power consumption, are still present. This contribution presents a solution for application-driven adaptation of our reconfigurable architecture at register transfer level (RTL) to reduce the resource requirements and power consumption while keeping the flexibility and performance for a predefined set of applications. Furthermore, implemented runtime adaptive features like online routing and configuration sequencing will be presented and discussed. A presentation of the prototype chip of this architecture designed in 90 nm standard cell technology manufactured by TSMC will conclude this contribution. 1. Introduction Reconfigurable architectures aim to reach the performance and energy-efficiency of application-specific integrated circuits while the flexibility is increased, therefore closing the gap between ASICs and general-purpose processors. For data-oriented applications an increase in performance compared to general-purpose processors can be reached by mapping operations to a possibly large set of functional units, which are working in parallel. In contrast to ASICs, their actual function and the interconnection between the units are not determined during design and manufacturing but may be changed at runtime to support a wider range of applications. For example, in a mesh-based architecture, a flexible communication network connects the functional units (FUs) on demand. Since the FUs are communicating directly by exchanging the intermediate results through the communication network, memory accesses for temporary data storage are avoided and memory bandwidth usage is reduced to a minimum. The overall data throughput is at maximum and very close to the ideal performance that can be reached by ASIC implementations. However, this approach is not without limitations. The increased flexibility comes at the cost of additional hardware. The flexible communication network for FUs requires a lot of multiplexers, communication lines, configuration registers, and additional logic to control the configuration mechanisms. Depending on the type of the reconfigurable approach (coarse-grained or fine-grained) the overhead of the configuration registers and control logic can be considerable. An example for this fact is given by field-programmable gate arrays (FPGAs) [1, 2], which require a
%U http://www.hindawi.com/journals/ijrc/2012/832531/