Giulia Guidi

Achieving a 2x speedup in distributed sparse matrix multiplication by exploiting symmetry could redefine performance benchmarks in large-scale machine learning applications.

Achieving up to 6.8x speedup in SGD without sacrificing accuracy could revolutionize distributed training efficiency on GPUs.

Ditching the standard two-pass approach to sparse matrix multiplication on GPUs can yield 2x speedups, thanks to a clever estimation technique that avoids costly symbolic computation.