Cross-layer optimization of QRD accelerators
Upasna Vishnoi, Tobias G. Noll · 2013
QR decomposition accelerators are attractive SoC components for many applications. A new linear QRD array is proposed based on the Givens-algorithm and CORDIC (Coordinate Rotate Digital Computer) rotations. In order to achieve high area- and energy-efficiency a systematic design space exploration has to be applied. Thereby the strong interactions between the architecture, micro-architecture and circuit level have to be considered. Based on architecture templates and pre-characterization of basic building blocks an algebraic cost-model is presented allowing for a cross-layer design space exploration with adequate accuracy in reasonable time. As a result from this quantitative optimization approach, a 16×16-QRD macro performing 30 million QRDs per second on 0.3 mm2area and 160 pJ per QRD is derived.