Applying the Roofline model for Deep Learning performance optimizations

This work addresses performance optimization for deep learning practitioners using Intel hardware, but it is incremental as it applies an existing model to a new system.

The paper tackles the problem of optimizing deep learning performance by automatically generating Roofline models for Non-Unified Memory Access systems, using Intel Xeon as an example, and evaluates the efficiency of deep learning primitives in the Intel oneDNN Library, showing performance improvements in specific benchmarks.