jCAE has been successfully tested on Intel Xeon servers with 16 GB RAM. Since there is no binary distribution of OpenCascade for this platform, we had to compile it from sources, which worked pretty smoothly. We then compiled occjava, our Java bindings for OpenCascade, and that is all, jar files do not have to be regenerated.
With 16GB RAM, we were able to mesh a single patch with more than 13 millions of triangles (this is not a hard limit, but it is better to keep some free memory to avoid degraded performance due to overuse of garbage collection).
And the icing on the cake is that this test demonstrated performance bottleneck of current algorithm on some shapes, and SVN version runs now much faster.