Machine Learning Frameworks Interoperability, Part 1: Memory Layouts and Memory Pools

Efficient pipeline design is crucial for data scientists. When composing complex end-to-end workflows, you may choose from a wide variety of building blocks, each of them specialized for a dedicated task. Unfortunately, repeatedly converting between data formats is an error-prone and performance-degrading endeavor. To learn more on framework interoperability, check out our presentation at NVIDIA's GTC 2021 Conference. Zero-copy functionality is a crucial technique to efficiently copy data across GPU-accelerated data science frameworks: TensorFlow, PyTorch, MXNet, cuDF, CuPy, Numba, and JAX (see Figure 2).