It is true that the SVM is relatively good at handling a large number of
features, especially for the linear caase with useless dimensions simply
getting rejected by being assigned a 0 weight in the regression.
However I do disagree at least to an extent with yur comment based on my
experience on my (Large) datasets: The presence of a large number of
irrelevant features does degrade the performance, especially so if the
kernel is not linear. If our objective is to attain good regression
performance with low errors, even though the features may not be noisy(in
measurement etc) we would still like to prune the irrelevant features for
the regression problem.