A network of memory and coherence controllers is provided which interconnected nodes in a cache-coherent multi-processor system. The nodes contain multiple processors operatively connected via respective caches to associated memory and coherence controllers. The system supports better processor utilization and better application performance by reducing the latency in accessing data by performing proactive speculative data transfers. In being proactive, the system speculates, without specific requests from the processors, as to what data transfers will reduce the latency and will make data transfers according to information derived from the system at any time that data transfers could be made.