One-sided Jacobi SVD on CUDA GPUs: Profiling-Guided Kernel Fusion and Data-Movement Optimization

Using end-to-end timing, Nsight Systems, and Nsight Compute, this paper shows that the baseline is limited less by the rotation formula or device throughput than by a mismatch between Jacobi rounds and CUDA Runtime boundaries.

The study evaluates a fused cooperative kernel across fourteen workload classes and checks numerical behavior through reconstruction error, singular values, effective-subspace orthogonality, and convergence traces. It demonstrates how profiling evidence can turn a measurable control-plane problem into a verifiable execution-structure optimization.