Fixed-point architectures manipulate data with relatively small word- lengths, which could offer the advantages of a significantly lower area, latency, and power consumption in embedded systems. However, fixed-point arithmetic introduces quantization noises which modify computation accuracy and, so, the application integrity. Our approach consists in studying fixed-point accuracy privileging analytical methods. This ones reduces drastically the conversion time compared to methods based on simulation during the floating-point arithmetic to fixed-point arithmetic conversion process of the application.
Our studies include :
- Efficient implementation of equalization and synchronous algorithms for digital applications
- Adaptive precision scaling: modifying precision according to channel and environment modifications
- Multi-constraints implementations of digital communication applications: the idea is to find a satisfying implementation according to different constraints (multi accuracy, and/or latency, and/or cost)
- Sampling frequency effects on computing accuracy
- Study of new arithmetics for digital applications (approximate computing, vectorial quantization…)