A new hardware implementation of the triangular neighborhood function (TNF) for ultra-low power, Kohonen self-organizing maps (SOM) realized in the CMOS 0.18žm technology is presented. Simulations carried out by means of the software model of the SOM show that even low signal resolution at the output of the TNF block of 3-6 bits (depending on input data set) does not lead to significant disturbance of the learning process of the neural network. On the other hand, the signal resolution has a dominant influence on the overall circuit complexity i.e. the chip area and the energy consumption. The proposed neighborhood mechanism is very fast. For an example neighborhood range of 15 a delay between the first and the last neighboring neuron does not exceed 20 ns. This in practice means that the adaptation process starts in all neighboring neurons almost at the same time. As a result, data rates of 10-20 MHz are achievable, independently on the number of neurons in the map. The proposed SOM dissipates the power in-between 100 mW and 1 W, depending on the number of neurons in the map. For the comparison, the same network realized on PC achieves in simulations data rates in-between 10 Hz and 1 kHz. Data rate is in this case linearly dependend on the number of neurons.
Przeanalizowano możliwości wykonywania obliczeń równoległych za pomocą instrukcji wielozadaniowych nowoczesnych procesorów sygnałowych (DSP) i zbadano ich przydatność do wybranych zadań przetwarzania sygnałów. Analizowano cechy architektur tych układów, przy czym szczególną uwagę poświęcono architekturze VLIW (Very Long Instruction Word). W celu zilustrowania skuteczności instrukcji wielozadaniowych, zbadano efektywność wykonywania typowych zadań przetwarzania sygnałów za pomocą stałoprzecinkowego procesora sygnałowego Blackfin (w zestawie SDK 2.01 ze środowiskiem VisualDSP++ 4.5) oraz za pomocą zmiennoprzecinkowych procesorów Sharc i Tiger-Sharc.
EN
In this paper possibilities of realization of parallel computations using multi-issue instructions of modern digital signal processors (DSPs) have been analyzed and their usefulness for chosen signal processing tasks has been investigated. Features of architectures of these processors have been analyzed with special attention paid to the VLIW (very long instruction word) architecture. In order to illustrate efficiency of multi-issue instructions, effectiveness of realization of typical signal processing tasks has been investigated using fixed-point Blackfin processor (by means of VisualDSP++ 4.5) environment with Software Development Kit 2.01) as well as using Sharc and TigerSharc floating-point processors.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.