Randomized SVD already allows this. Typically the calculation of PCA on a data set around the size that they use (1-100TB) is I/O bound. I work in genomics wher

Randomized SVD already allows this. Typically the calculation of PCA on a data set around the size that they use (1-100TB) is I/O bound. I work in genomics where this is a common problem. Their solution is interesting, but I'm not really sure if it would be more efficient due to parallelism. Better accuracy is important, but it's also unclear how much that affects the major PCs.