Future SoCs will feature embedded FPGAs (eFPGAs) to enable flexible and efficient implementations of high-throughput digital signal processing applications. Current research projects on and emerging products containing FPGAs are mainly based on "standard FPGA"-architectures that are optimised for a very wide range of applications. The implementation costs of these FPGAs are dominated by a very complex interconnect network. This paper presents a method to improve the efficiency of eFPGAs by tailoring them for a certain application domain using a parametrisable architecture template derived from the results of a systematic evaluation of the requirements of the application domain.

Two different architectures are discussed, a reference architecture to illustrate the methodology and possible optimisation measures as well as a specialised arithmetic-oriented eFPGA for applications like correlators, decoders, and filters. For the arithmetic-oriented architecture, a novel logic element (LE) and a special interconnect architecture that was designed with respect to the connectivity characteristics of regular datapaths, are presented. For both architecture templates, physically optimised implementations based on an automatic design approach have been created.

As a first cost comparison of these implementations with standard FPGAs, the LE-density (number of logic elements per mm²) is evaluated. For the arithmetic-oriented architecture, the LE-density could be increased by an order of magnitude compared to standard architectures.