Looking at thin matrix SVD, it appears much faster than everyone else. I’m curious what it’s doing differently at a high level and if there’s any tradeoff in accuracy. I also wonder how it compares to MKL, which is typically the winner in all these benchmarks on Intel.