Architecture is driven by the demands of application computing requirements and enabled by advancing technology. Architecture both structures and balances resources to deliver functionality and performance. The Architecture Working Group investigated the largely unexplored regime of parallel computer architecture at the far reaches of PetaFLOPS performance. Although orders-of-magnitude beyond contemporary experience, key issues of technology, parallelism, latency, bandwidth, size, and cost were examined to determine potential approaches and their respective feasibility. Surprises and uncertainty characterized the results, and in the process some popular assumptions were discarded. In the end, a PetaFLOPS computing system was conceived-but its gestation is still uncertain, strongly influenced by exigencies poorly understood and even less well controlled. The results of this inquiry reveal a path, milestones, and decision points that can be used to guide planners and establish early research directions.