I think there still is room for improvement in speed of t-SNE. One way would be using threads while computing joint probabilities. But main thing to do is profiling and checking for which parts are actually causing the slowdown.
Pharo is not sequential, but process are green threads all running on the same VM thread. Maybe in the future the VM will manage multi-threads map them to the core but this is not the case right now. The only solution is to run multiple-images.