Most Digital Signal Processing (DSP) courses concentrate on algorithm development in floating-point arithmetic (e.g. Matlab), but 90% of DSP processors sold do not offer hardware support for floating-point, or even fractions, only integers. Fixed-point algorithms allow smaller data sizes and more efficient processing in both processors and FPGA implementations, but only if the programmer is careful to avoid the pitfalls of limited precision. This talk will include a comparison of the performance of fixed-point and floating-point implementations, a description of the speaker’s efforts to create tutorial materials on fixed-point implementation, and a short tutorial on the four kinds of error introduced by fixed-point arithmetic.

Host: Behnaam Aazhang