Poster 96: TSQR on TensorCores

Authors: Hiroyuki Ootomo (Tokyo Institute of Technology), Rio Yokota (Tokyo Institute of Technology)

Abstract: Tall-Skinny QR (TSQR) is an efficient algorithm for calculating the QR decomposition of m x n matrices where m << n, which is done by recursively performing QR decomposition on subdivided blocks of the tall and skinny matrix. Such operations are useful for low-rank approximation methods, which are replacing more and more dense linear algebra in both scientific computing and machine learning fields. The present work focuses on the implementation of this important algorithm on Tensor Cores, which are available on the latest NVIDIA GPUs. We evaluate the speed, accuracy, and stability of TSQR on TensorCores.

Best Poster Finalist (BP): yes

Poster: PDF
Poster summary: PDF


Back to Poster Archive Listing