We have 24 display stations, each connected to a Benq 7765PA DLP projector, and each of these projects a 1024x768 image, which is 1/24th of the total image, i.e., a 6K x 3K pixel display on a 14' x 7' screen. Each display station has a 1.6GHz Pentium 4 processor and a Matrox G450 card, chosen for fast pixel movement, not rendering. Each display station is connected via Gigabit Ethernet to 9 render nodes. We have 240 render nodes in all, but we use a 9-for-1 fan-in instead of 10-for-1 because we get better network performance, and we can then use 24 of the nodes as geometry generators. Each render node has a 1.6GHz Pentium 4 processor and an Nvidia GeForce4 TI 4400 graphics card. We're using a modified version of Stanford's Chromium package to do the distributed rendering. The full pipeline looks like:
control node -24-for-1-fan-out-> geometry nodes -9-for-1-fan-out-> render nodes -1-for-9-fan-in-> display nodes
We get 37 fps from the system on simple geometry, e.g., Mark Kilgard's "Atlantis". On more taxing applications, such as our own reconstruction (from point cloud data) of the Confederate submarine, the H.L. Hunley,
Alignment of the 24 projectors is
totally in software. Calibration points are projected, read back
with a camera, and then used to build the inverse of each projection transformation.
This compensates for physical alignment errors without any need for expensive,
pan-tilt-rotate mounts.
