KDD Cup: Crossroads

A lot of papers claim SVM works well in high-d. In the examples they provide, high-d is roughly in the tens of thousands. What if you have a dataset that has 200,000 dimensions? The paper "A Divisive Information-Theoretic Feature Clustering" claims SVMs are problematic in high-d. It is however, not a well cited paper (only 5 citations) and uses Linear SVMs. Could it also be that the issue was with the software they used?