Why is GPU not available for classification?

The speedup in GPU computing relies on performing many operations onto the same piece of data. This is an optimal situation for direct space alignment, where the same template is rotated many times and compared to a given particle. In PCA, the situation is different. The most numerically demanding part is the construction of the cross correlation matrix. Here, a particle read from the disk just experiences a rotation, a fourier filtering and a direct space scalar product. This workload is not enough to compensate for the overhead of transferring the particles into the GPU memory (which is normally the bottleneck in generic GPU computing!).