Runtime Usage#
The MemryX dataflow architecture is designed to achieve high-throughput inference performance. Since it is a “pipelined dataflow”–different from traditional CPU/GPU architectures–proper use of the MemryX Runtime is essential to achieve optimal performance when integrating into user applications.
Under the hood, there is also the Linux and Windows drivers (installed in Get Started), but these are not directly interacted with by the user.
Instead, the Python and C++ APIs provide a high-level interface to the MemryX hardware, allowing users to focus on their application logic rather than low-level hardware details.
This section provides deeper details on:
How to define and use async Callback functions with the MemryX Runtime.
Multiple Streams (e.g. multiple cameras) and how to use them with the Runtime in your application.
How to connect cropped pre- and post-processing models in your application, and commentary on performance.
Overview of the MXA-Manager daemon, and Shared and Local modes: two ways to control application access to MXA hardware.
How to use multiple accelerator modules on the same system, with automatic load balancing.
Using multiple DFPs at the same time on the same device, within applications or across multiple running processes.
See also
Docker: Using MemryX Runtime in Docker containers.
Tutorials: Example applications using the complete MemryX SDK.
Install Runtime (Linux): Installation instructions for the MemryX Runtime.
Architecture Overview: Overview of the MemryX architecture and dataflow model.