DaCe AD

Unifying high-performance automatic differentiation for ML and scientific computing

DaCe AD: Unifying High-Performance Automatic Differentiation for ML and Scientific Computing

An efficient automatic differentiation framework that, on average, is over 92× faster than JAX on gradient computation for scientific computing patterns.

Benchmarked on the NPBench suite. Unifies AD across ML and scientific computing workloads in a single framework.

  • Role: Lead author
  • Methods: DaCe, automatic differentiation, NPBench
  • Status: Published, CLUSTER 2025
  • Paper: arXiv:2509.02197

References