Abstract: Advanced collaborative display systems allow users to view a computing 'desktop' environment in a platform and location independent fashion. For real-time considerations, these systems become computationally very challenging, especially when video-streaming is included. The addition of stereoscopic video streaming is desirable in Virtual Environments (VE) created for the teaching of anatomy and surgery with real-time collaborative audio and video interactions at many locations. However this stresses the real-time requirements to the point at which realistic video is difficult to assure. For such a system to work, it is imperative that timing data be collected, analyzed, and understood. In this paper, we describe an experimental system designed primarily for collection of timing data required for robust collaborative medical training applications. The networked stereoscopic system uses a server-swappable multicast network protocol to stream real-time manipulations of 3D Virtual Body Structures (VBS, [1]) at the server site to all clients participating in the multicast session. The three visual modes (mono, psa, and asi) have dynamic tuning parameters for adjusting the parallax and the focal point for the rendered scene, allowing users to define individual stereoscopic comfort zones. Our main result demonstrates that optimizing the graphics module is critical in achieving the necessary rendering rates. Different techniques of utilizing various memory resources increased the number of polygons rendered per second by over seven million. Depending on the type of memory used the number of polygons rendered per second varies from 2.25 to 9.12 million.