Faster SVD via Accelerated Newton-Schulz Iteration
Askar Tsyganov ⋅ Uliana Parkina ⋅ Ekaterina Grishina ⋅ Sergey Samsonov ⋅ Maxim Rakhuba
Abstract
Traditional SVD algorithms rely heavily on QR factorizations, which scale poorly on GPUs. We show how the recently proposed Chebyshev-Accelerated Newton-Schulz (CANS) iteration can replace them and produce an SVD routine that is faster across a range of matrix types and precisions.
Successful Page Load