Using a cluster of commodity Alpha processors we compare two software platforms based on Linux and Windows NT and intended to support intensive scientic computations. Networking and compiler performance are separately analysed and then results for NAS parallel benchmarks are given. We nd that a compiler able to make good use of the cache is more importantthanlow network latency in obtaining high performance. We argue that for all types of cluster, the choice of compiler is critical in selecting a cost effective platform for computationally intensive scientic applications.