In this paper a new type of parallel computer architecture dedicated to real-time vision is proposed. The proposed architecture is based upon the Beowulf concept — i.e. COTS computing nodes interconnected by a high-speed network and running standard, free software tools — but also takes into account the constraints of embedded vision systems, such as "on the fly" processing of video streams, small volume, and low power consumption. Its main originality lies in the presence of two separate communication media: a dedicated bus for fast video i/o and a standard, switched inter-connection network for inter-process communication. Another distinctive point is the use of a high-level parallel programming model, based upon algorithmic skeletons. In order to assess the validity of the proposal, a first prototype has been built and is described in the paper. It uses G4 motherboards coming from Apple Cube mass market computers, a Fast Ethernet communication network and a FireWire (IEEE-1394) video bus. Results of preliminary benchmarks are presented.