Throughput, speed and cost considerations

An example has been described of the coding of a speech recognition algorithm onto a floating point digital signal processor by means of a high level language compiler. It has been shown that this can considerably reduce the engineering effort required, compared with that for a fixed point DSP using assembly language. It must be concluded however, that it is necessary to use assembly coding for computationally intensive routines if real-time performance is to be achieved. When devices are used in production volumes, the extra cost of floating point over fixed point devices will not be justified by the savings in engineering costs. However, if the cost of engineering effort involved in programming the DSP is significant, when compared with the total cost of the DSP devices, then the extra cost for a floating point device can be justified by the saving of incremental engineering costs. This is particularly the case if further modifications of the algorithm may be required.