Registered Data

[00693] Enzyme: Fast and Effective Automatic Differentiation for Academia and Industry

  • Session Time & Room : 3C (Aug.23, 13:20-15:00) @E605
  • Type : Industrial Contributed Talk
  • Abstract : Automatic differentiation (AD) is key to training neural networks, Bayesian inference, and scientific computing. Applying these techniques requires rewriting code in a machine learning framework or manually providing derivatives. We present Enzyme, an AD extension for the industry-standard LLVM/MLIR compiler. Enzyme differentiates programs in any LLVM-based language. Unlike traditional tools, Enzyme performs AD on optimized code, resulting in a 4.2x speedup on the CPU and orders of magnitude speedup on the GPU.
  • Classification : 65Kxx, 65Yxx, 68Vxx
  • Format : Talk at Waseda University
  • Author(s) :
    • William Steven Moses (MIT)
    • Valentin Churavy (MIT)
    • Ludger Paehler (TUM)
    • Oleksandr Zinenko (Google)