SparseDitto speeds GPU kernels 2.79x

SparseDitto’s LLM agents generated workload-specific sparse GPU kernels that delivered a 2.79x geometric-mean speedup over cuSPARSE on an NVIDIA H200. The paper reports a 78.5x maximum on H200 and up to 3.39x faster full-batch graph-neural-network training, after planning, coding, verifying, and measuring several designs for each matrix and operator. For teams building the verified engineering loops described in our analysis of NVIDIA’s agent toolkit, the result is a reason to benchmark generated kernels against vendor libraries when data shape is stable—but not to generalize an author-reported maximum across workloads.