Design Space Exploration for ST’s Neural Compilation toolchain


Fabrizio Indirli, Cristina Silvano, Giuseppe Desoli, Andrea Ornstein
Polytechnic of Milan, STMicroelectronics

Presentation title

Design Space Exploration for ST’s Neural Compilation toolchain

Authors

Fabrizio Indirli
Polytechnic of Milan
Cristina Silvano
Polytechnic of Milan
Giuseppe Desoli
STMicroelectronics
Andrea Ornstein
STMicroelectronics

Presentation type

Technical presentation

Abstract

To enable efficient NN inference on the edge, embedded Neural Processing Units (NPUs) are becoming common in MCUs and SoCs: an example is STMicroelectronics experimental NPU and its associated compilation toolchain. This mapping tool exposes various compile-time and design-time parameters that can be configured, such as the nodes' decomposition strategy or the number of the NPU's computing units. Finding the optimal strategy to efficiently map a NN on the target hardware accelerator is a challenging task: given the vastness of this design space, automatic optimization techniques are needed. In this work, we integrated STMicroelectronics NPU compilation toolchain with an automatic exploration framework (MOST), and we compared several Design Space Exploration (DSE) techniques such as Simulated Annealing, Greedy Search and Genetic Algorithms to identify the most suitable methodology to drive the autotuning of the NPU compiler co-design parameters. We then applied these optimizations to select the best set of parameters to run the inference of the Tiny-Yolo DCNN using ST NPU.


Additional material

  • Presentation slides: [pdf]