The performance and accuracy of quantitative precipitation forecasts (QPF) based on NWP models have continuously increased in recent years, especially for light and moderate precipitation amounts, although high amounts and rare events still are difficult to handle. This constitutes a severe limitation for any operational activities due to the strong dependency on resolution, data assimilation and predictability. Nevertheless, the advantages associated with even a marginal anticipation and enhancement of early warnings for flash floods are often so valuable that the setting up of integrated atmospheric-hydrological forecasting systems is becoming well justified. One system of this kind was set up for the Arno river basin, Italy, which is quite peculiar as regards flood risk issues. Most of the territory is prone to frequent flood events, with high-risk levels due to the vulnerability of the unique artistic and cultural heritage, as demonstrated during the 1966 Florence flooding. The hydrology of the region is such that the rainfall-discharge delay varies from almost immediate to a day at most. The city of Florence itself is characterized by a rainfall-discharge delay between 8 and 12 hours. Flood forecasting in such a complex environment needs to rely on a variety of monitoring and prediction tools, from rainfall-runoff modeling to rainfall forecasts, real time control of soil moisture and hydro-meteorological monitoring. Furthermore, these tools need to be used in a flexible way in order to constitute the optimal prediction chain for the event at hand. In this work the improvement of Quantitative Precipitation Forecasts produced by the Regional Atmospheric Modeling System (RAMS) for severe historical rainstorms and floods is evaluated statistically against rain gauge observations and the accuracy of flood forecasts, with a view to demonstrating the impact of: 1)horizontal and vertical model resolution; 2)initialization times (forecast lead times); 3)assimilation of sea surface temperatures; 4)assimilation of spatially continuous satellite estimated precipitation (diabatic initialization). Furthermore, the effectiveness and performance of the low cost, highly scalable parallel computing system used for the experimentation as well as for the operational setting as well as its possible further enhancements are discussed.